bullshitai.dev

RESEARCH

Agent memory: findings from my experiments

How does memory for my agents actually work? A week with Hindsight, a plain SQLite memory built with my agents, and what a 15-question comparison measured.

METHOD
My agent threads, frozen run artifacts and a read-only diagnosis. No new runs for this article.
WINDOW
5–11 Oct 2026. Fifteen questions, one frozen comparison.
RESULT
13 of the 14 points of the gap came from data my memory never indexed. Retrieval and packing showed two mechanisms worth changing.
my memory's score on 15 questions; Hindsight scored 23 / 30
9 / 30
points of the gap from questions it had no data for
13 of 14
questions where its keyword search returned anything
0 of 15
where search put the right message; the 900-character excerpt stopped before the answer
rank 2

The question: what does an agent's memory actually do?

I run Hermes, my self-hosted agent setup, as daily infrastructure. Several agent profiles talk to me over Discord and Telegram and run cron jobs. Each profile keeps its own conversation history in SQLite.

Memory is the layer that should let an agent answer “what did we decide about this last week?” without me pasting the thread again. I wanted to know what that layer really does between a conversation and an answer: what it stores, how it finds things, and what ends up in the model’s context. Until this week the layer was Hindsight, an open-source memory server, and I wasn’t happy with what it recalled. So I ran two experiments. First I tuned Hindsight and watched how it behaved on my setup. Then I built a simpler memory of my own with my agents and compared the two on 15 questions about my own history.

Most of the hands-on work was done by agents: Hermes itself and the Claude Code workers it launched. I made the calls and read the reports. Some of those reports overstated a result and were corrected later; where that happened, I say so.

The short version: my homemade memory scored 9/30 against Hindsight’s 23/30. Tracing every question showed that most of that gap measured the test, not retrieval. Underneath it, the trace found two mechanisms in my own code, one in retrieval and one in context packing.

From conversation to answer: the two routes I testedHindsight as its docs describe it and as configured on my setup · hermes-mem as tested (0.3 candidate)Hindsightstores what an LLM extractsConversationretained every10 turnsRetain (LLM)writes facts andobservationsRecall4 searches,fused, rerankedBudgettop 60 reranked,512-token blockThe modelgets the memoryonly within 8 shermes-memstores the messages themselvesConversationtool output andnoise droppedStoremessages, BGE-M3vectors, cardsSearchkeyword (strictAND) + semanticExcerptsmessage heads,900 charactersThe modelgets the excerptsthat fit the budgetHindsight stages from its docs; budget, block size and the 8-second Hermes deadline from my config (7 October).hermes-mem: the tested 0.3 candidate. The public 0.2.0 returns whole episodes and keeps reviewed records instead of cards.
From conversation to answerthe two routes I testedHindsightstores LLM extractsConversationretained every10 turnsRetain (LLM)writes facts andobservationsRecall4 searches,fused, rerankedBudgettop 60 reranked,512-token blockThe modelgets the memoryonly within 8 shermes-memstores the messagesConversationtool output andnoise droppedStoremessages, BGE-M3vectors, cardsSearchkeyword (strictAND) + semanticExcerptsmessage heads,900 charactersThe modelgets the excerptsthat fit the budgetHindsight: its docs plus my config (7 Oct).hermes-mem: the tested 0.3 candidate;public 0.2.0 returns whole episodes.
Two routes, not a ranking. Hindsight’s stages are from its docs; the numbers are my settings. hermes-mem is the tested 0.3 candidate.
Text version of the figure

Hindsight (stages from its docs; settings from my config on 7 October)

  1. Conversation, retained every 10 turns.
  2. Retain: an LLM writes facts and observations. What is stored is extracted text.
  3. Recall: semantic, BM25 keyword, graph and temporal search, fused by reciprocal rank and reranked.
  4. Budget: the low budget passes the top 60 candidates to the reranker; the result fills a 512-token memory block.
  5. The model gets the memory only if it arrives within Hermes’ 8-second wait.

hermes-mem (the tested 0.3 candidate)

  1. Conversation, with tool output and other noise dropped.
  2. Store: the messages themselves, BGE-M3 vectors and a few agent-written cards.
  3. Search: keyword (strict AND) and semantic.
  4. Excerpts: the head of each message, cut to 900 characters.
  5. The model gets the excerpts that fit the budget.

The public 0.2.0 returns whole episodes instead of single messages and keeps reviewed records instead of cards.

Hindsight in use: extraction, recall and the deadline

Hindsight has three verbs. Retain uses an LLM to turn conversations into facts, experiences and consolidated “observations”. Recall searches them: semantic, BM25 keyword, graph and temporal search, fused by reciprocal rank, reranked, then trimmed to a token budget. Reflect runs a slower reasoning loop over the memories. Memories live in isolated banks. All of that is from Hindsight’s own docs; what follows is what I observed on my setup.

My Hermes plugin config, as an agent printed it on 7 October: automatic recall on, asynchronous, observation-only, the low budget, a 512-token memory block, a 20-second timeout. Retain ran every ten turns. Reranking went through Luna, an LLM reached through my ChatGPT subscription. The embedding model was E5-small. My agents never recorded the Hindsight version or the retain model in those threads, so I can’t give you either.

What I saw falls into three layers.

What goes in (our setup). We fed cron output into retain, and one cron job alone produced 2,526 of 7,786 consolidated records, 32.4%. The low budget passes the top 60 candidates to the reranker. In one audit, 19 of 36 records we had agreed were useful were cut before the reranker ever saw them. Raising that cap to 250 made reranking take 11–28 seconds instead of about 8. The experiment also used up a local 100-calls-a-day reranker limit that production shared, so for a while my live agents ran without reranking. Several test runs stopped on our own orchestration: broken task contracts, a supervisor relaunch loop, a cron job that couldn’t report back.

How it reaches a turn (the integration). Hermes waits at most 8 seconds for memory before a turn. In one screening run, recall missed that deadline in 99 of 177 successful requests. One answer came back from Hindsight in 8.48 seconds and Hermes dropped it. In asynchronous mode, a turn gets the memory that was fetched for the previous turn.

How extraction treated my data. In an audit of stored memories, two of my agent’s recommendations had been saved as my decisions, and it happened again later in live context. Retain rules were ours too, so I count that as mixed. Long Ukrainian questions went over the 500-token query limit and got an HTTP 400. Reflect was too slow to run every turn.

Then the tuning. Doubling the memory block from 512 to 1,024 tokens looked like a win on 20 tuning questions: 4 fully supported answers against 2. On held-out questions, both settings fully supported 1 of 9. The agent’s own verdict that night was that the settings we tested did not solve the problem. On 10 October I stopped the Hindsight experiments and moved to the second one.

A plain memory to compare against: hermes-mem

The second experiment was deliberately plain: take the messages, drop tool output and other noise, add an embedding to each one so the agent can search by meaning, and keep a few large cards about me, my intentions and the assistant itself. The agent writes those cards with a tool when it notices something worth keeping, or when I ask it to. That became hermes-mem: SQLite, FTS5 for keywords, and BGE-M3 for embeddings.

The model choice came from 22 synthetic Ukrainian and English queries run on a Mac.

Three embedding models on 22 synthetic queriesUkrainian and English, run on a Mac. It picked a default; it says nothing about my real history.Right dialogue ranked first, of 22Median warm query, ms (linear)0510152004008001,200BGE-M3 (INT8)21/2252 msQwen3-Embedding-0.6B21/221,065 msMiniLM17/2213 msSame 22 queries for all three; one host. Not a measurement on my real history.
Three embedding models22 synthetic UK/EN queries, on a MacRight dialogue ranked first, of 2201020BGE-M3 (INT8)21/22Qwen3-Embedding-0.6B21/22MiniLM17/22Median warm query, ms (linear)06001,200BGE-M3 (INT8)52 msQwen3-Embedding-0.6B1,065 msMiniLM13 msA small synthetic set; it picked a default.
22 synthetic queries, one host. Both axes are linear. It picked a default; it doesn’t tell you how any of these models do on my real history.
Text version of the figure
ModelRight dialogue ranked first (of 22)Median warm query
BGE-M3 (INT8)2152 ms
Qwen3-Embedding-0.6B211,065 ms
MiniLM1713 ms

BGE-M3 matched Qwen3’s 21 of 22 at a fraction of its query time. On eight extra queries with mixed-language wording, BGE-M3 got 8 of 8 and the other two 7 of 8. That is a small synthetic set.

Two infrastructure incidents from that week tend to get retold as model results, so here is what they were. The first Mac run stopped because Docker hit the macOS keychain before the container even started. Separately, an E5-small benchmark on my VPS was OOM-killed at 1,536 MiB while the model was still loading, before it had embedded a single text. The 41 seconds in that log is the time to the crash, not embedding throughput. Neither incident says anything about embedding quality.

Then the default profile’s history went in. Its database held 171,270 rows in 443 sessions, but rows aren’t messages: 74,115 were tool output, 55,861 were empty assistant turns, and plenty were copies left behind by context compaction. After filtering there were 2,834 messages, forming 1,288 episodes and 3,567 chunks to embed.

The first pilot embedded 148 of those chunks, about 4%. At roughly four seconds a chunk on the VPS, the rest would have taken about four hours. The agent reported the profile as “activated”. Search worked; the backfill hadn’t run and automatic sync was off. I switched my default profile to hermes-mem and back again the same evening.

Next we tried to carry over what Hindsight knew. Its export held 20,320 memory units, 8,107 of them observations, built from 654 source documents. Only 1,420 observations made it into a test database. Not because the rest were worthless:

Where 8,107 Hindsight observations went in the importAccess filtering, not quality scores · observations only (8,107 of the export's 20,320 memory units)01,0002,0003,0004,0005,0006,0007,000Imported into the test database1,420Conversation belongs to another profile6,216No link to a conversation252No owner tag140From scheduler sessions79ObservationsThe five groups add up to 8,107, so each observation is counted once. Not imported is not the same as worthless.
Where 8,107 observations wentAccess filtering, not quality scores02,0004,0006,000Imported into the test database1,420Conversation belongs to another profile6,216No link to a conversation252No owner tag140From scheduler sessions79ObservationsObservations only: 8,107 of 20,320 memory units.The groups add up to 8,107; each is counted once.
Access filtering, not quality scores. The import kept observations by where they came from and who could see them. The five groups add up to 8,107.
Text version of the figure
Import outcomeObservations
Imported into the test database1,420
Conversation belongs to another profile6,216
No link to a conversation252
No owner tag140
From scheduler sessions79
All observations in the export8,107

The import filtered by access rules, and I had first read the number as a value judgement. Once that was clear, we started the design over with a smaller core.

The minimum version kept only messages, embeddings and cards. Review caught a privacy bug before acceptance: calls without a session fell into one shared session:unknown bucket. A second fix kept source references stable when a message was rolled back. The accepted revision passed 81 tests on each of two Python versions, 31 of 31 integration checks and 43 of 43 install checks. It is an unpushed 0.3 candidate. The public repository still holds 0.2.0, which returns whole episodes instead of single messages and keeps reviewed records instead of cards.

A 15-question comparison

The comparison used 15 questions about my own history, each worth 2 points under a written rubric. The candidate searched the default profile only: 2,914 messages, 4,384 of 4,384 chunks embedded, nine cards. Hindsight answered from its usual bank.

The result: my candidate 9/30, Hindsight 23/30. Hindsight did better on 10 questions, mine on 1, and 4 were ties.

The agent’s report did say the corpora weren’t equal. I read the headline as a verdict on my retrieval anyway, and that reading was the next thing to test.

One thing to keep straight before reading further: 9/30 is graded answer points. It is not recall@k, and it doesn’t tell you how often search found the right message.

What the score measured

So I went through it question by question: is the answer not in the database, or is it there and search didn’t find it? A separate agent traced each question through every stage, from source to corpus, search, context, answer and grade. It changed nothing and ran no new answers or judges.

Where the 21 points wentPoints lost · one frozen run, 15 questionsTest setup: the answer was only in data it never hadIn hermes-mem's own pipeline05101520Missing source data9 questions18Context clipped1 question2Incomplete answer1 question1Points lost (2 per question, 30 possible)
Where the 21 points wentPoints lost · one frozen runTest setup: answer only in data it never hadIn hermes-mem's own pipeline05101520Missing source data · 9 questions18Context clipped · 1 question2Incomplete answer · 1 question1Points lost (2 per question)
One frozen run, not a ranking. 15 questions, 2 points each. The hermes-mem 0.3 candidate scored 9 of 30, so 21 points went unscored; four questions got full marks. The long bar is the test setup: those answers were only in the two agent profiles the candidate never indexed. Images: wide PNG · phone PNG.
Text version of the chart
What happenedKindQuestionsPoints lost
Missing source dataThe answer was only in the two agent profiles it never indexedTest setup918
Context clippedFound the right message at rank 2, then cut the answer off at 900 charactershermes-mem packing12
Incomplete answerHad the facts in context, left part of the answer out (Hindsight lost the same point)Answer step11
Full-mark answersAnswered in full40
Total1521

The 14-point gap to Hindsight (23 − 9) splits by the same causes: 13 corpus + 2 packing + 0 answer − 1 where mine scored higher. An attribution of one frozen run, not a re-score. Also test setup: the answer key was written from Hindsight's own memories, and the test sent no reworded queries. Also a retrieval finding: hybrid keyword search found nothing on all 15 questions.

The corpus decided most of it. Hindsight’s bank mixes the history of three agent profiles. My candidate indexed one. In 9 of the 15 questions, every fact in the expected answer came from the two profiles the candidate never saw. Those nine questions account for 13 of the 14 points of the gap. In nine of my candidate’s ten zero-point answers, it honestly said the notes didn’t have it. The tenth was a real invention: it pulled in an unrelated detail.

The answer key leaned one way. The expected answers were written from Hindsight’s own observations. That bias didn’t come from leakage; everything was frozen in the right order. It came from using one contender’s memories as the answer key. In one question the key said I had decided to add a table to a document. The source conversation holds my request for a review and the agent’s recommendation to add that table. There is no decision from me. My candidate answered that no decision was recorded, which is true, and scored zero.

The 14-point gap by primary cause
Primary causeQuestionsPoints my candidate lostPoints of the gap
Answer not in its corpus91813
Retrieved, then cut from the context122
Support present, answer incomplete110
Full-mark answers40−1
Total152114

The table attributes one frozen result to causes. It doesn’t claim what an equal test would score.

Retrieval and packing: two mechanisms in my code

Inside the data it did have, my own code limited what reached the model in two ways.

Hybrid search was semantic-only. In hybrid mode the keyword arm requires every word of the question to match, first as words, then as trigrams. Relaxing to OR is switched off there. On natural-language questions that never fires: no keyword hits on any of the 15 original questions or on 28 of 30 reworded variants. The strictness was deliberate. In the project’s first evaluation, OR matches on common words flooded the fusion and cut hybrid hit@1 from 0.93 to 0.14. It still left hybrid search running on one arm. Where the answer was in the corpus, semantic search usually found it: rank 1 or 2 for three questions. One more sat at rank 15 for the original wording and moved to rank 2 with a short Ukrainian rewording. The tool accepts such rewordings from the agent. The test harness passed none.

The right message, the wrong passage. Search returns the head of each message, cut to 900 characters; 59 of the 120 results returned in the run were cut. In the table question above, search put the right message at rank 2, but the table starts near character 4,435.

Found the message, missed the passageOne case from the trace: the table question. Search ranked this message 2nd (a rank, not a position).The message, by character positioncontinues; total length not recorded01,0002,0003,0004,0005,0006,000table starts near 4,435Excerpt of 900 characters, the default: table left outExcerpt of 4,000, the configured maximum: table left outExcerpt of 6,000, tried in the rerun: table includedbut then only 3 results fit the budget instead of 8Character position within the message
Found the message, missed the passageOne case: search ranked this message 2ndThe message, by characterlength not recorded02,0004,0006,000table near 4,435Excerpt: 900, the defaulttable left outExcerpt: 4,000, the maximumtable left outExcerpt: 6,000, in the reruntable in; 3 results fit, not 8Character position
One observed case, not a retrieval rate. From a one-variable rerun of the packing step. Rank and character position are different units; the message’s total length wasn’t recorded.
Text version of the figure
  • Search ranked the right message 2nd.
  • Inside that message, the table starts near character 4,435.
  • An excerpt of 900 characters (the default) left the table out.
  • An excerpt of 4,000 characters (the configured maximum) also left it out.
  • An excerpt of 6,000 characters included the table, but then only three results fit the budget instead of eight.

My first idea was to centre the excerpt on the best-matching chunk. The trace ruled it out. The best-matching chunk of that message was the first one (similarity 0.619); the chunk holding the table scored 0.562. Search found the right message for the wrong part of it, and the plugin has no tool to read a whole message.

Both behaviours are in the public 0.2.0 on GitHub as well as the unpushed candidate, as of 11 October. Neither has been changed yet.

What to test next

Same data on both sides, first. Either index all three profiles for my candidate or restrict Hindsight to the one profile. Until that runs, I don’t know how much of the 13 points would close. Some facts could live in places neither side looked.

An answer key that doesn’t come from one contender’s memory. Write it from the original conversations, and check who actually said what.

Score the stages separately. Is the answer in the corpus? At what rank did search put it? Did it survive packing? Was the answer right? Graded points alone collapse all four into one number.

Agent status reports as claims to check. “Activated” meant search worked on 4% of the chunks. “45%” wasn’t accuracy. The reports were usually honest a message later, which is a fine trait in a colleague and a bad one in a dashboard.

The next experiments for hermes-mem are proposals, not shipped changes: return the matching passages instead of message heads, or add a bounded tool that reads a whole message; let the keyword arm fall back to OR or BM25 at a lower weight; pass the agent’s rewordings in the test the way the product intends. Another open-source memory tool, Basic Memory, documents both of the first two ideas: a matched passage next to the head preview and a full-note read tool. In my table case only the full read would have helped.

None of this crowns Hindsight or rules it out. It beat my candidate on a test that leaned its way, and on my setup its integration had real costs: the 8-second deadline, mixed-up authorship, a reranker cap that cut useful records. Those are separate questions, and this week answered none of them cleanly.

Method, sources and limits

Sources. Two private Discord threads between me and my agents (5–11 October 2026), the frozen comparison artifacts, and a read-only diagnosis that replayed the accepted search code with the pinned BGE-M3 model. Specific numbers come from tool output or reports in those threads; I kept a private ledger linking each one to its message and time. Raw transcripts, memory contents and private paths are not published.

What was not done. No equal-corpus rerun. The diagnosis didn’t open the other two profiles’ databases, so “not in the corpus” is bounded to the one profile and Hindsight’s export. The lookups used to find each answer were written after the fact to locate support, not to measure search. The exports are incomplete history; a missing detail isn’t proof that it never happened. Hindsight’s version and retain model weren’t recorded in the threads I used.

Public sources. Hindsight README and docs for retain, recall and reflect; hermes-mem (public 0.2.0 at 587a162; the tested 0.3 candidate 56120d6 is local); SQLite FTS5; the BGE-M3 model card; Basic Memory’s search docs. Vendor docs describe claimed behaviour, not my measurements.

The chart. Drawn for this article as SVG: a plain bar chart of where the 21 unscored points of the frozen run went, by cause. Its frame, axis grid, legend and direct labels borrow the grammar of the charts on Anthropic’s Claude Opus 5.5 page. The text version under the chart carries the same content.

← things/
Banner image credit and licence
Banner image: Twemoji "Brain" emoji (U+1F9E0), by Twitter, Inc and other
contributors, from the Twemoji project. Used unmodified.

Copyright (c) 2014–2021 Twitter
Copyright (c) 2022–present Jason Sofonia & Justine De Caires
(the copyright lines of the Twemoji project; this file has not changed
upstream since 2019)

Source:  https://github.com/jdecked/twemoji/blob/b6b55fef1e8636b540a6d016a4729ca8cdf2e60b/assets/svg/1f9e0.svg
Project: https://github.com/jdecked/twemoji

Graphics licence: Creative Commons Attribution 4.0 International (CC BY 4.0),
https://creativecommons.org/licenses/by/4.0/
No changes were made to the artwork; it is only shown at a larger size.
The material is provided as-is, without warranties (sections 5 and 6 below).

The full licence text follows, as published in the project's LICENSE-GRAPHICS.

------------------------------------------------------------------------------

Attribution 4.0 International

=======================================================================

Creative Commons Corporation ("Creative Commons") is not a law firm and
does not provide legal services or legal advice. Distribution of
Creative Commons public licenses does not create a lawyer-client or
other relationship. Creative Commons makes its licenses and related
information available on an "as-is" basis. Creative Commons gives no
warranties regarding its licenses, any material licensed under their
terms and conditions, or any related information. Creative Commons
disclaims all liability for damages resulting from their use to the
fullest extent possible.

Using Creative Commons Public Licenses

Creative Commons public licenses provide a standard set of terms and
conditions that creators and other rights holders may use to share
original works of authorship and other material subject to copyright
and certain other rights specified in the public license below. The
following considerations are for informational purposes only, are not
exhaustive, and do not form part of our licenses.

     Considerations for licensors: Our public licenses are
     intended for use by those authorized to give the public
     permission to use material in ways otherwise restricted by
     copyright and certain other rights. Our licenses are
     irrevocable. Licensors should read and understand the terms
     and conditions of the license they choose before applying it.
     Licensors should also secure all rights necessary before
     applying our licenses so that the public can reuse the
     material as expected. Licensors should clearly mark any
     material not subject to the license. This includes other CC-
     licensed material, or material used under an exception or
     limitation to copyright. More considerations for licensors:
	wiki.creativecommons.org/Considerations_for_licensors

     Considerations for the public: By using one of our public
     licenses, a licensor grants the public permission to use the
     licensed material under specified terms and conditions. If
     the licensor's permission is not necessary for any reason--for
     example, because of any applicable exception or limitation to
     copyright--then that use is not regulated by the license. Our
     licenses grant only permissions under copyright and certain
     other rights that a licensor has authority to grant. Use of
     the licensed material may still be restricted for other
     reasons, including because others have copyright or other
     rights in the material. A licensor may make special requests,
     such as asking that all changes be marked or described.
     Although not required by our licenses, you are encouraged to
     respect those requests where reasonable. More_considerations
     for the public: 
	wiki.creativecommons.org/Considerations_for_licensees

=======================================================================

Creative Commons Attribution 4.0 International Public License

By exercising the Licensed Rights (defined below), You accept and agree
to be bound by the terms and conditions of this Creative Commons
Attribution 4.0 International Public License ("Public License"). To the
extent this Public License may be interpreted as a contract, You are
granted the Licensed Rights in consideration of Your acceptance of
these terms and conditions, and the Licensor grants You such rights in
consideration of benefits the Licensor receives from making the
Licensed Material available under these terms and conditions.


Section 1 -- Definitions.

  a. Adapted Material means material subject to Copyright and Similar
     Rights that is derived from or based upon the Licensed Material
     and in which the Licensed Material is translated, altered,
     arranged, transformed, or otherwise modified in a manner requiring
     permission under the Copyright and Similar Rights held by the
     Licensor. For purposes of this Public License, where the Licensed
     Material is a musical work, performance, or sound recording,
     Adapted Material is always produced where the Licensed Material is
     synched in timed relation with a moving image.

  b. Adapter's License means the license You apply to Your Copyright
     and Similar Rights in Your contributions to Adapted Material in
     accordance with the terms and conditions of this Public License.

  c. Copyright and Similar Rights means copyright and/or similar rights
     closely related to copyright including, without limitation,
     performance, broadcast, sound recording, and Sui Generis Database
     Rights, without regard to how the rights are labeled or
     categorized. For purposes of this Public License, the rights
     specified in Section 2(b)(1)-(2) are not Copyright and Similar
     Rights.

  d. Effective Technological Measures means those measures that, in the
     absence of proper authority, may not be circumvented under laws
     fulfilling obligations under Article 11 of the WIPO Copyright
     Treaty adopted on December 20, 1996, and/or similar international
     agreements.

  e. Exceptions and Limitations means fair use, fair dealing, and/or
     any other exception or limitation to Copyright and Similar Rights
     that applies to Your use of the Licensed Material.

  f. Licensed Material means the artistic or literary work, database,
     or other material to which the Licensor applied this Public
     License.

  g. Licensed Rights means the rights granted to You subject to the
     terms and conditions of this Public License, which are limited to
     all Copyright and Similar Rights that apply to Your use of the
     Licensed Material and that the Licensor has authority to license.

  h. Licensor means the individual(s) or entity(ies) granting rights
     under this Public License.

  i. Share means to provide material to the public by any means or
     process that requires permission under the Licensed Rights, such
     as reproduction, public display, public performance, distribution,
     dissemination, communication, or importation, and to make material
     available to the public including in ways that members of the
     public may access the material from a place and at a time
     individually chosen by them.

  j. Sui Generis Database Rights means rights other than copyright
     resulting from Directive 96/9/EC of the European Parliament and of
     the Council of 11 March 1996 on the legal protection of databases,
     as amended and/or succeeded, as well as other essentially
     equivalent rights anywhere in the world.

  k. You means the individual or entity exercising the Licensed Rights
     under this Public License. Your has a corresponding meaning.


Section 2 -- Scope.

  a. License grant.

       1. Subject to the terms and conditions of this Public License,
          the Licensor hereby grants You a worldwide, royalty-free,
          non-sublicensable, non-exclusive, irrevocable license to
          exercise the Licensed Rights in the Licensed Material to:

            a. reproduce and Share the Licensed Material, in whole or
               in part; and

            b. produce, reproduce, and Share Adapted Material.

       2. Exceptions and Limitations. For the avoidance of doubt, where
          Exceptions and Limitations apply to Your use, this Public
          License does not apply, and You do not need to comply with
          its terms and conditions.

       3. Term. The term of this Public License is specified in Section
          6(a).

       4. Media and formats; technical modifications allowed. The
          Licensor authorizes You to exercise the Licensed Rights in
          all media and formats whether now known or hereafter created,
          and to make technical modifications necessary to do so. The
          Licensor waives and/or agrees not to assert any right or
          authority to forbid You from making technical modifications
          necessary to exercise the Licensed Rights, including
          technical modifications necessary to circumvent Effective
          Technological Measures. For purposes of this Public License,
          simply making modifications authorized by this Section 2(a)
          (4) never produces Adapted Material.

       5. Downstream recipients.

            a. Offer from the Licensor -- Licensed Material. Every
               recipient of the Licensed Material automatically
               receives an offer from the Licensor to exercise the
               Licensed Rights under the terms and conditions of this
               Public License.

            b. No downstream restrictions. You may not offer or impose
               any additional or different terms or conditions on, or
               apply any Effective Technological Measures to, the
               Licensed Material if doing so restricts exercise of the
               Licensed Rights by any recipient of the Licensed
               Material.

       6. No endorsement. Nothing in this Public License constitutes or
          may be construed as permission to assert or imply that You
          are, or that Your use of the Licensed Material is, connected
          with, or sponsored, endorsed, or granted official status by,
          the Licensor or others designated to receive attribution as
          provided in Section 3(a)(1)(A)(i).

  b. Other rights.

       1. Moral rights, such as the right of integrity, are not
          licensed under this Public License, nor are publicity,
          privacy, and/or other similar personality rights; however, to
          the extent possible, the Licensor waives and/or agrees not to
          assert any such rights held by the Licensor to the limited
          extent necessary to allow You to exercise the Licensed
          Rights, but not otherwise.

       2. Patent and trademark rights are not licensed under this
          Public License.

       3. To the extent possible, the Licensor waives any right to
          collect royalties from You for the exercise of the Licensed
          Rights, whether directly or through a collecting society
          under any voluntary or waivable statutory or compulsory
          licensing scheme. In all other cases the Licensor expressly
          reserves any right to collect such royalties.


Section 3 -- License Conditions.

Your exercise of the Licensed Rights is expressly made subject to the
following conditions.

  a. Attribution.

       1. If You Share the Licensed Material (including in modified
          form), You must:

            a. retain the following if it is supplied by the Licensor
               with the Licensed Material:

                 i. identification of the creator(s) of the Licensed
                    Material and any others designated to receive
                    attribution, in any reasonable manner requested by
                    the Licensor (including by pseudonym if
                    designated);

                ii. a copyright notice;

               iii. a notice that refers to this Public License;

                iv. a notice that refers to the disclaimer of
                    warranties;

                 v. a URI or hyperlink to the Licensed Material to the
                    extent reasonably practicable;

            b. indicate if You modified the Licensed Material and
               retain an indication of any previous modifications; and

            c. indicate the Licensed Material is licensed under this
               Public License, and include the text of, or the URI or
               hyperlink to, this Public License.

       2. You may satisfy the conditions in Section 3(a)(1) in any
          reasonable manner based on the medium, means, and context in
          which You Share the Licensed Material. For example, it may be
          reasonable to satisfy the conditions by providing a URI or
          hyperlink to a resource that includes the required
          information.

       3. If requested by the Licensor, You must remove any of the
          information required by Section 3(a)(1)(A) to the extent
          reasonably practicable.

       4. If You Share Adapted Material You produce, the Adapter's
          License You apply must not prevent recipients of the Adapted
          Material from complying with this Public License.


Section 4 -- Sui Generis Database Rights.

Where the Licensed Rights include Sui Generis Database Rights that
apply to Your use of the Licensed Material:

  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
     to extract, reuse, reproduce, and Share all or a substantial
     portion of the contents of the database;

  b. if You include all or a substantial portion of the database
     contents in a database in which You have Sui Generis Database
     Rights, then the database in which You have Sui Generis Database
     Rights (but not its individual contents) is Adapted Material; and

  c. You must comply with the conditions in Section 3(a) if You Share
     all or a substantial portion of the contents of the database.

For the avoidance of doubt, this Section 4 supplements and does not
replace Your obligations under this Public License where the Licensed
Rights include other Copyright and Similar Rights.


Section 5 -- Disclaimer of Warranties and Limitation of Liability.

  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.

  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.

  c. The disclaimer of warranties and limitation of liability provided
     above shall be interpreted in a manner that, to the extent
     possible, most closely approximates an absolute disclaimer and
     waiver of all liability.


Section 6 -- Term and Termination.

  a. This Public License applies for the term of the Copyright and
     Similar Rights licensed here. However, if You fail to comply with
     this Public License, then Your rights under this Public License
     terminate automatically.

  b. Where Your right to use the Licensed Material has terminated under
     Section 6(a), it reinstates:

       1. automatically as of the date the violation is cured, provided
          it is cured within 30 days of Your discovery of the
          violation; or

       2. upon express reinstatement by the Licensor.

     For the avoidance of doubt, this Section 6(b) does not affect any
     right the Licensor may have to seek remedies for Your violations
     of this Public License.

  c. For the avoidance of doubt, the Licensor may also offer the
     Licensed Material under separate terms or conditions or stop
     distributing the Licensed Material at any time; however, doing so
     will not terminate this Public License.

  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
     License.


Section 7 -- Other Terms and Conditions.

  a. The Licensor shall not be bound by any additional or different
     terms or conditions communicated by You unless expressly agreed.

  b. Any arrangements, understandings, or agreements regarding the
     Licensed Material not stated herein are separate from and
     independent of the terms and conditions of this Public License.


Section 8 -- Interpretation.

  a. For the avoidance of doubt, this Public License does not, and
     shall not be interpreted to, reduce, limit, restrict, or impose
     conditions on any use of the Licensed Material that could lawfully
     be made without permission under this Public License.

  b. To the extent possible, if any provision of this Public License is
     deemed unenforceable, it shall be automatically reformed to the
     minimum extent necessary to make it enforceable. If the provision
     cannot be reformed, it shall be severed from this Public License
     without affecting the enforceability of the remaining terms and
     conditions.

  c. No term or condition of this Public License will be waived and no
     failure to comply consented to unless expressly agreed to by the
     Licensor.

  d. Nothing in this Public License constitutes or may be interpreted
     as a limitation upon, or waiver of, any privileges and immunities
     that apply to the Licensor or You, including from the legal
     processes of any jurisdiction or authority.


=======================================================================

Creative Commons is not a party to its public licenses.
Notwithstanding, Creative Commons may elect to apply one of its public
licenses to material it publishes and in those instances will be
considered the "Licensor." Except for the limited purpose of indicating
that material is shared under a Creative Commons public license or as
otherwise permitted by the Creative Commons policies published at
creativecommons.org/policies, Creative Commons does not authorize the
use of the trademark "Creative Commons" or any other trademark or logo
of Creative Commons without its prior written consent including,
without limitation, in connection with any unauthorized modifications
to any of its public licenses or any other arrangements,
understandings, or agreements concerning use of licensed material. For
the avoidance of doubt, this paragraph does not form part of the public
licenses.

Creative Commons may be contacted at creativecommons.org.