Agent Learning · Checks for the lessons an agent takes from its own experience

Captured, tested, applied, and kept.

An agent learns from experience when every failure, correction, incident and surprise it goes through is captured as a lesson record with its source episode, date and context, each lesson is distilled into a general statement with the situation it applies in and what to do differently, it becomes a standing rule only after a test on enough episodes, later decisions follow it and name it, it stays in the lesson store from session to session with every retirement recorded, it carries into new tasks inside its stated scope and stays out of the rest, new learning is checked against the lessons already verified before it takes effect, and the report the principal reads states what was captured, validated, adopted, applied and retained with the records behind each figure. Agent Learning checks each of those with eight tools: eight checks. Your agent sends the episodes and lesson records, the lesson statements, the evidence and tests, the later decisions, the lesson store snapshots, the new tasks, the updates or the learning report in the call, and each tool returns a signed report that names what is uncaptured, still raw, adopted on an anecdote, shelved, decaying, overreaching, overwritten or inflated, with the fix for each. The tools read the recorded material; the lessons and the work stay with the agent and its principal.

See the tools Read the contract

Tool access: 0.10 USDC / call
Tools: Eight tools: eight checks
Connection: /mcp
Result: Signed report

The eight tools.

Each tool reads one stage in the life of a lesson an agent takes from its own experience and answers one question about it. Every tool reads the recorded material in the call: episodes, lesson records, statements and logs, evidence and tests, later decisions, store snapshots and changes, scopes and new tasks, verified lessons and updates, and learning reports go in, and a classification comes out. The report names the episode, lesson, decision, snapshot, task, update or figure behind every finding, with the fix for each gap, so your agents and you see why it reads the way it does.

`wever_learning-capture-check`

Reads: The experience episodes an agent went through, each with its kind and log, read for lesson wording such as next time or root cause, with the lesson records captured from them.
The signed report states: Episodes that carry a lesson with no lesson record, records that state no lesson, records with no source episode, date or context, and records written long after the experience.
Price: 0.10 USDC per call
Verdicts: experience_not_capturedcapture_incompletecapture_informalexperience_captured

`wever_learning-distillation-check`

Reads: Each lesson as stated, the episodes and raw log it came from, the situation it applies in and what to do differently there.
The signed report states: Lessons with no statement of their own, statements that repeat their log, statements tied to one event such as yesterday or a ticket number, and lessons with no situation or action.
Price: 0.10 USDC per call
Verdicts: lessons_still_rawdistillation_partialno_distillationlessons_distilled

`wever_learning-validation-check`

Reads: Each candidate and adopted lesson with its supporting and contradicting episodes, its test against evidence and the dates it was tested and adopted.
The signed report states: Lessons adopted as standing rules on a single episode, against as many contradicting episodes, or before their test, and lessons adopted on thin evidence or untested.
Price: 0.10 USDC per call
Verdicts: anecdote_as_rulevalidation_lightvalidation_absentlessons_validated

`wever_learning-application-check`

Reads: Each adopted lesson with the later decisions it covers, whether each decision followed it, named it, or recorded why it went the other way.
The signed report states: Later decisions that went against an adopted lesson with no recorded reason, decisions that applied a lesson without naming it, and the overrides with their reasons.
Price: 0.10 USDC per call
Verdicts: lessons_shelvedapplication_partialno_applicationlessons_applied

`wever_learning-retention-check`

Reads: Snapshots of the lesson store as the agent loaded it session after session, with the record of lessons added, retired and merged.
The signed report states: Lessons that dropped out of the store with no recorded retirement, lessons that dropped out and came back, unrecorded additions, removals with no reason and unversioned snapshots.
Price: 0.10 USDC per call
Verdicts: lessons_decayingretention_unevenretention_absentlessons_retained

`wever_learning-transfer-check`

Reads: Each lesson with where it holds and where it does not, and the new tasks it was considered for, inside or outside that scope.
The signed report states: Lessons carried into tasks outside their stated scope, tasks inside the scope that went without the lesson, lessons with no stated scope and lessons worded always or never with no limits.
Price: 0.10 USDC per call
Verdicts: transfer_overreachingtransfer_partialno_transfertransfer_scoped

`wever_learning-regression-check`

Reads: The lessons the agent already verified, and each update since with whether it was checked, what it contradicts, when that was detected and how it was settled.
The signed report states: Contradictions with verified lessons that were never detected, verified lessons overwritten or superseded on weaker evidence, contradictions left on hold and contradictions found after the fact.
Price: 0.10 USDC per call
Verdicts: lessons_overwrittenregression_containedregression_uncheckedno_regression

`wever_learning-report-check`

Reads: The learning report written for the principal: the lessons captured, validated, adopted, applied and retained as stated, whether each figure links to its records, and the record behind it.
The signed report states: Figures above the record, figures with no records behind them, claims such as learns from every mistake against a record of lessons not applied or kept, and figures left out or understated.
Price: 0.10 USDC per call
Verdicts: report_inflatedreport_incompletereport_missinglearning_report_sound

Start with the two free tools.

On https://weverlabs.com/mcp-free, anonymous and with no wallet, two Agent Learning tools show what a sound learning loop holds and what a well-kept learning journal looks like beside a flawed one. Each returns an unsigned result in the family's report shape and ends with a line naming the main tools.

`wever_learning-snapshot`

Reads: An agent name and a kind of agent, such as a support agent, a coding agent or a multi-agent team.
Returns: The orientation checklist of a sound learning loop for that kind of agent, part by part from captured experience to the principal report, with the items the kind of agent adds and the check that reads each part.
Access: Free on /mcp-free

`wever_learning-journal-sample`

Reads: Nothing, or one sample journal: a support desk, a coding assistant or a research assistant.
Returns: One worked learning journal kept right beside a flawed one that leaves an episode uncaptured, keeps lessons as raw logs, adopts an anecdote untested, ignores a lesson, lets one drop out silently, carries one past its scope and lets new learning overwrite a verified lesson, each piece annotated with why it passes or fails, scored by seven of the family checks. The checks read your own journal.
Access: Free on /mcp-free

How agents and people use it.

Your agents

Pull the checklist with wever_learning-snapshot when setting up a learning loop, so capture, testing, retention and regression checks are in place from the first lesson.
Call the check that answers the question on https://weverlabs.com/mcp at the moment it matters: after an episode that went wrong, when a lesson is written up, before a lesson becomes a standing rule, when later work is reviewed, at the start of each session, before a lesson is carried into new work, before new learning takes effect and before the report reaches the principal.
Act on the verdict and keep the signed report beside the lesson it covers, so every adoption, retirement and update carries the check behind it.

You

See which lessons your agent has taken from its own experience, what each rests on, and where each one applies.
Spot a single bad day turning into a standing rule while it is still a candidate, and see a lesson that quietly dropped out of the store.
Read a learning report where every figure links to its records, and verify any check against Wever's published trust key whenever you need to show where a verdict came from.

A worked example

One claim, and a new standing rule.

The request

{
  "name": "wever_learning-validation-check",
  "arguments": {
    "lessons": [
      {
        "id": "l1",
        "lesson": "Confirm the card on file before refunding an order placed before a card change",
        "status": "adopted",
        "adopted_at": "2026-11-12",
        "evidence": 5,
        "against": 0,
        "tested": true,
        "tested_at": "2026-11-11"
      },
      {
        "id": "l2",
        "lesson": "Offer store credit first on damaged-item claims",
        "status": "adopted",
        "adopted_at": "2026-11-14",
        "evidence": 1,
        "against": 0,
        "tested": false
      },
      {
        "id": "l3",
        "lesson": "Quote the standard window when the order shows no shipping zone",
        "status": "adopted",
        "adopted_at": "2026-11-15",
        "evidence": 2,
        "against": 2,
        "tested": true,
        "tested_at": "2026-11-16"
      }
    ]
  }
}

What the report returns

Verdict: anecdote_as_rule.

Finding, anecdote adopted	1 lesson was adopted as a standing rule on a single episode or none: l2 (1 supporting episode).
Finding, contradicted adopted	1 lesson was adopted with as many contradicting episodes as supporting ones: l3 (2 for, 2 against).
Finding, adopted before test	1 lesson was adopted before the test that should have validated it: l3 (adopted 2026-11-15, tested 2026-11-16).
Finding, adopted untested	1 lesson was adopted without a test against evidence: l2.
Fix, anecdote adopted	Keep a lesson resting on one episode as a candidate until enough episodes support it.
Fix, contradicted adopted	Retire or rescope an adopted lesson that as many episodes contradict as support.
Fix, adopted before test	Test a lesson against evidence before adopting it, not after.
Fix, adopted untested	Test every lesson against evidence, such as a replay or a held-out set of episodes, before adopting it.

Anecdotes have become standing rules: anecdote adopted, contradicted adopted and adopted before test.

The store credit lesson rests on one damaged-item claim and was adopted untested, so a single case now decides every claim like it. The shipping window lesson was adopted a day before its test, with as many episodes against it as for it. The card lesson was tested first, on five refunds, and reads as validated. Moving the store credit lesson back to a candidate and retesting the shipping lesson keeps both open to the next claims that come in.

Call it from your agent.

Connect to https://weverlabs.com/mcp and call any of the eight tools. Each call costs 0.10 USDC on Base. Eligible calls share 10 free calls per wallet over a rolling 30 days with the other paid tools. For free access, send X-Wever-Wallet-Proof signed over the exact JSON arguments and the tool route. The payment challenge supplies the x402 terms for a paid call.

Every report comes with an Ed25519 receipt signed with Wever's published trust key. The Agent Learning contract lists the schemas, limits, verdicts and signing details. For whether outside feedback was incorporated, Agent Feedback; for stored knowledge and its retrieval, Agent Knowledge; for how memory is structured, the memory design check on Agent Oversight, and for what memory keeps, Agent Memory; for the confidence an agent states, Agent Uncertainty; for the promises it makes, Agent Commitments.

Browse all tools