Two technicians inspect a machine on a production line

Amplify your experts’ judgement.

Metis is an open-source toolkit for capturing fragments of expert practice and making them available to AI agents as memory, with human review and agreed conditions for use.

See how it works Read the paper

01 / What it is

Memory for what experts know but cannot fully explain.

Metis (μῆτις) takes its name from the Greek idea of practical intelligence: knowing how to act in a particular situation.

Tacit knowledge is what we know but find difficult to explain. Metis helps record a specific part of that practice: what an expert noticed, how they responded and the circumstances of that response. We call this record a tacit fragment.

Each fragment keeps its source and conditions for use attached. After human review, it can become part of a fourth layer of agent memory, alongside procedures, facts and past events.

Proceduralwhat should happen
Semanticwhat is known
Episodicwhat happened
Tacitwhat experience adds

Tacit memory

Reviewed experience helps an agent apply its other knowledge to a particular situation.

02 / Why Metis

The procedure doesn’t explain every decision.

An experienced operator hears a pump change note under high load. A quality specialist spots a batch that looks wrong. Their judgement draws on experience and the situation as well as the written procedure.

Procedures describe what should happen; logs record what happened. What can be missing is why the expert responded differently, which cues mattered and when the same response would be wrong. An agent needs that context to use the account as guidance.

Work as imagined

Reduce load only when the alarm threshold is crossed. SOP-17

Work as done

Ease back earlier, when high load meets a dull sound. An experienced operator, night shift

Work as imagined

Reduce load only when the alarm threshold is crossed.

SOP-17

Work as done

Ease back earlier, when high load meets a dull sound.

An experienced operator, night shift

Work as imagined

Pass the batch unless the lab measure flags drift.

The release procedure

Sample

Lab result

Pending

Work as done

Flag a change in surface sheen before the lab result arrives.

A quality specialist, finishing area

Reference example

This batch

Work as imagined

A completed handover form means the handover is complete.

The handover procedure

Handover form✓✓✓

Form complete

Work as done

Ask the incoming lead to confirm open threads out loud.

A team lead, night to day

Outgoing lead

Open threads

Incoming lead

Verbal confirmation needed

Metis records what the expert noticed, its context and their response, so the account can be reviewed before an agent uses it.

An atlas of know-how

Expertise takes many forms.

These examples draw on just a few of the seventeen kinds of expertise explored in the paper. Use the atlas to find those that matter in your work and see how they could be captured and reviewed.

Explore the atlas — 17 kinds of know-how

Select a segment to see an example and how it could be captured.

Procedural and embodied · K1

Procedural

Routines already written down. The baseline the other kinds diverge from.

Capture approach: Document ingestion; procedure analysis
In practice: SOP-17 itself.

Procedural and embodied · K2

Embodied

Skilled bodily action: the feel of a tool starting to bind.

Capture approach: Multimodal observation; expert confirmation
In practice: A fitter easing a fastener by feel.

Procedural and embodied · K3

Rhythmic

Timing and tempo: how long to wait, when to move.

Capture approach: Action-timing analysis; pause and tempo comparison
In practice: A nurse holding a step a beat longer than the protocol asks.

Material and equipment · K4

Equipment-specific

The quirks of one machine that the manual for its type never mentions.

Capture approach: Maintenance logs; cross-equipment comparison
In practice: This pump runs warm after a weekend stop.

Material and equipment · K5

Material

The feel and behaviour of a material lot.

Capture approach: Outcome correlation with material lots; worker narration
In practice: Slowing a mix because this resin sets faster.

Material and equipment · K6

Tool-extended

Skill that lives in the pairing of a person and a tool.

Capture approach: Tool-use traces; substitution events
In practice: Same operator, different probe, different reading.

Perceptual and aesthetic · K7

Sensory

Hearing a pump change note before any alarm. Cues learned through years of exposure.

Capture approach: Multimodal cue capture; on-cue narration
In practice: An operator hears a dull note from a pump under high load and eases it back before the alarm.

Perceptual and aesthetic · K8

Aesthetic

Quality distinctions learned through exemplars: a batch that looks off.

Capture approach: Expert annotation of exemplars
In practice: A quality specialist flags a resin batch after comparing its surface sheen with reference examples.

Inferential · K9

Heuristic

Rules of thumb for recognising and responding to exceptions.

Capture approach: In-flow prompt; exception logging
In practice: Re-running a passing sample because one peak looks odd.

Inferential · K10

Diagnostic

Reasoning from symptom to cause under uncertainty.

Capture approach: Critical Decision Method; incident reconstruction
In practice: Interpreting vibration and a dull sound under high load as possible signs of bearing trouble.

Inferential · K11

Anticipatory

Recognising signs of what may happen next.

Capture approach: Pre-event prompting on detected divergence
In practice: Easing back before the alarm.

Meta-cognitive and affective · K12

Meta-cognitive

Knowing the edge of your own competence: when to pause, when to ask.

Capture approach: Reflective probing; help-seeking pattern analysis
In practice: The lead who stops a handover for one more question.

Meta-cognitive and affective · K13

Affective-regulatory

Composure under pressure, and the signals that go with it.

Capture approach: Consented affective cues; post-event reflection
In practice: Consented discussion of cues during work or in later reflection.

Social and normative · K14

Collaborative

Handoffs, dependencies and coordination between people.

Capture approach: Handoff analysis; cross-actor coordination
In practice: A team lead asks the incoming lead to confirm open threads aloud, even though the handover form is complete.

Social and normative · K15

Cultural-narrative

The stories that carry local norms.

Capture approach: Long-form interview; story collection
In practice: Explore through stories, interviews and shared interpretation.

Social and normative · K16

Judgemental-ethical

Weighing what is right when the rule and the situation pull apart.

Capture approach: Post-event walk-through; structured reflection
In practice: People retain responsibility for the ethical judgement.

Social and normative · K17

Strategic

Sense-making across levels: what matters here, this quarter.

Capture approach: Strategic interview; cross-level alignment review
In practice: The supervisor who reads a plant target and knows what it means for this line.

03 / How it works

How it works.

An expert notices something and acts on it. If a capture agent spots a difference between the recorded action and the written procedure, it asks what prompted the decision. The expert checks the account, and reviewers agree where and how it can be used. An AI agent can then draw on that experience when the situation fits, within the agreed limits.

Illustrated examples

Pump vibration

Operator · Pump A · high load · night shift

Observe

Start with a workflow record.

A connected workplace system supplies a record of what happened. The capture agent compares it with the procedure and the available context.

What the workflow system recorded

Load was reduced before the alarm threshold.

Operator · Pump A · high load · night shift

The action is recorded. The reason is not.

Depth where it matters.

Short questions during work bring cues and changes to light. Focused interviews explore the experience, options and judgements behind them.

Discovery

Understand the practice

Refresh

Notice changes in the work

Follow-up

Investigate an unclear account

Discovery builds an understanding of the work. Refresh and follow-up keep it current.

About the capture methods

Some decisions need a focused interview. The capture approach combines short questions during work with cognitive task analysis: interviews that examine the cues, options and judgements behind a decision.

Discovery

Understand the practice

A Knowledge Audit maps the expertise. Critical Decision Method interviews examine difficult decisions in detail.

Refresh

Notice changes in the work

Short questions help identify new cues, recurring adjustments and changes since the original account.

Follow-up

Investigate an unclear account

A focused follow-up interview revisits a new situation, conflicting evidence or an account that needs clarification.

Built on CHAP.

CHAP is a protocol for collaboration between people and AI agents. In the Metis reference flow, it records the practitioner’s confirmation, human review and why guidance was returned or withheld, so those decisions can be traced.

Explore CHAP Inspect the pump example’s recorded sequence

Discuss your workflow

04 / For builders

Run Metis locally.

Run the pump example, inspect its records, then connect Metis to your application. The demo uses supplied observations and needs no model.

Install and run the example.

Start in a Python 3.10+ virtual environment. Run the terminal commands, or switch to Python for an API example.

python -m pip install "metis-memory==0.1.2"
metis demo manufacturing-pump-vibration

# Inspect the records created by the demo.
metis fragment list
metis memory list
metis audit read

from metis import MetisEngine
from metis.conditions.context import TacitContext
from metis.consent.model import ConsentRecord, ConsentStatus

engine = MetisEngine()
engine.join_default_participants()

# Supplied account, consent and review decision for this example.
fragment = engine.capture_observation(
    {
        "observation_id": "OBS-1",
        "work_as_done": "Ease back early on a dull cue.",
        "context": TacitContext(
            equipment_family="centrifugal_pump",
            operating_mode="high_load",
        ),
    },
    consent=ConsentRecord(consent_status=ConsentStatus.granted),
    category="K7_sensory",
).fragment  # Evidence: not yet usable by an agent.

# In production, connect authenticated human review.
engine.tier2_review(fragment.fragment_id, "promoted_to_advisory")

matching = TacitContext(
    equipment_family="centrifugal_pump",
    operating_mode="high_load",
    risk_class="moderate",
)
different = TacitContext(equipment_family="gear_pump")

print(len(engine.retrieve(matching).eligible))
print(engine.retrieve(different).blocked[0].reason)

From the pump demo

Fragment: One account promoted to Advisory.
Guidance: Available when the context matches.
Collaboration record: 40 entries tracing the decisions.

metis-memory 0.1.2 · Python 3.10+ · Apache-2.0

GitHub Python package Documentation Ethical use

Inspect the records.

See how a fragment records an observation, its conditions and permitted use, and trace the decisions made during confirmation and review. The records below come from the supplied example.

Inside a tacit fragment

Six fields keep the observation with its source, conditions, evidence, permission and review. The record captures one part of the practitioner’s experience.

Pump vibration · TF-00001

Experienced operators reduce throughput earlier when high-load operation coincides with low-frequency vibration and a dull acoustic cue.

Human-confirmedContext-boundRevocable

Six-field schema

The cue and the response

Schema field: content

A description of what the practitioner does, which they confirm as faithful. One useful part of a practice, with the limits of that account.

cue: dull acoustic note and low-frequency vibration under high load
response: reduce throughput earlier than the procedure requires

Where it came from

Schema field: provenance

The observation, originating practitioner, capture method and review lineage remain attached. Attribution and consent are explicit.

observed by: human:operator@plant_a
captured by: observe, infer, whisper, confirm
consent: granted; visible to the worker; withdrawable

Where it applies

Schema field: conditions

Applicability is written as context the retrieval gate can check. A missing required value does not count as a match.

equipment: centrifugal pump, PUMP-A, line 3
context: high load, night shift, before an alarm
excludes: start-up

How strong the evidence is

Schema field: confidence

Recurrence, outcomes, counterexamples and uncertainty support review. A confidence score alone does not grant permission to use the fragment.

recurrence: four cases in the supplied fixture
outcome: three alarm events avoided
strength: moderate

What it may do

Schema field: authority

Advisory status permits conditional decision support. The originating operator’s action does not become an automatic action for the agent.

layer: advisory
limits: no automatic throughput reduction; confirm the cue with the operator; escalate high risk

Who checked it

Schema field: validation-state

Tier 1 checks descriptive fidelity. Tier 2 assesses relevance, evidence and normative alignment before granting a defined use.

tier 1: description confirmed by the practitioner
tier 2: promoted to Advisory by the Mission Group
state: subject to review, expiry, supersession and revocation

The evidence behind the pump example

This is the sequence from the original supplied example. The choices you make in the illustration above do not rewrite this record.

01A Capture Cell opens as a CHAP workspace.Coordinator
Sequence
000
Event
workspace.create
Actor
service:coordinator
02The operator joins.Operator
Sequence
001
Event
participant.join
Actor
human:operator@plant_a
03The bounded whisperer agent joins.Whisperer agent
Sequence
002
Event
participant.join
Actor
agent:whisperer#v1
04The human reviewers join as a group.Human review group
Sequence
003
Event
participant.join
Actor
group:mission-group
05The capture task opens and observation OBS-1 is recorded as an artefact.Operator
Sequence
005 to 007
Event
task.create / complete
Actor
human:operator@plant_a
06Candidate IC-0001 inferred. A hypothesis, never trusted knowledge.Whisperer agent
Sequence
008 to 009
Event
task.create / complete
Actor
agent:whisperer#v1
07"You noticed something before the formal measure changed. What cue made you pause?"Whisperer agent
Sequence
010
Event
whisper.ask
Actor
agent:whisperer#v1
08The operator confirms. Fidelity only.Operator
Sequence
013
Event
whisper.answer
Actor
human:operator@plant_a
09TF-00001 stored in the Evidence layer with provenance, conditions and consent.Operator
Sequence
014 to 021
Event
task.create / complete
Actor
human:operator@plant_a
10Submitted for review: relevance, alignment, risk.Human review group
Sequence
022
Event
review.request
Actor
group:mission-group
11Promoted to Advisory by human decision.Human review group
Sequence
023
Event
decide.approve
Actor
group:mission-group
12Memory object TM-00001 created with its use constraints.Human review group
Sequence
024 to 029
Event
task.create / complete
Actor
group:mission-group
13Retrieval allowed on the matching pump, blocked on a different one, agent context assembled.Assistant agent
Sequence
030 to 039
Event
task.create / complete
Actor
agent:assistant#v1

Hash links make alterations detectable. A decision record preserves accountability; its existence does not establish that the decision was correct.

Inspect the structured record

The working record expands the six-part contract with consent, lineage, review dates, expiry and revocation. These values come from the supplied demonstration, not a production deployment.

{
  "fragment_id": "TF-00001",
  "title": "Early throughput reduction on dull acoustic cue",
  "content": "Experienced operators reduce throughput earlier when high-load operation coincides with low-frequency vibration and a dull acoustic cue.",
  "category": "K7_sensory",
  "domain": "perceptual_aesthetic",
  "source_pathway": "exogenous",
  "provenance": {
    "observed_by": "human:operator@plant_a",
    "originating_participant": "human:operator@plant_a",
    "capture_cell": "wsp_pump_vibration",
    "source_pathway": "exogenous",
    "source_event": "OBS-1",
    "source_logs": [],
    "source_artefacts": [
      "art_01HF7YAT02000KRVQKEBZ99Y1A",
      "art_01HF7YAT040017HQF6WQYJKW2M",
      "art_01HF7YAT06001VAK6TB3XVXT3Y"
    ],
    "timestamp": "2026-08-18T00:54:29.696648+00:00",
    "capture_method": "observe->infer->whisper->confirm (synthetic)",
    "human_confirmed_by": "human:operator@plant_a",
    "mission_group_reviewed_by": "group:[email protected]",
    "model_provider": "ollama",
    "model_name": "gemma4",
    "model_prompt_template": null,
    "model_input_refs": [],
    "model_output_ref": null,
    "model_output_status": "draft_pending_human_review",
    "human_review_status": "tier1_confirmed",
    "model_assist_refs": [
      "MA-0001",
      "MA-0002"
    ]
  },
  "conditions": {
    "site": "plant_a",
    "area": "utilities",
    "line": "line_3",
    "equipment_family": "centrifugal_pump",
    "equipment_id": "PUMP-A",
    "product_family": null,
    "material_lot": null,
    "operating_mode": "high_load",
    "shift_pattern": "night",
    "role": "operator",
    "risk_class": null,
    "trigger_context": "pre_alarm",
    "environmental_conditions": {},
    "exclusion_conditions": [
      {
        "operating_mode": "startup"
      }
    ],
    "valid_from": null,
    "valid_until": null
  },
  "evidence": {
    "recurrence_count": 4,
    "supporting_cases": [
      "case-1"
    ],
    "comparison_baseline": "SOP-17",
    "outcome_link": "avoided 3 alarm events",
    "uncertainty": null,
    "counterexamples": [],
    "review_notes": [],
    "evidence_strength": "moderate"
  },
  "confidence": 0.3,
  "authority_layer": "advisory",
  "validation_state": "promoted_to_advisory",
  "consent": {
    "consent_required": true,
    "consent_status": "granted",
    "attribution_mode": "role",
    "visibility": "agent_visible",
    "withdrawal_allowed": true,
    "withdrawal_constraints": null,
    "worker_visible_record": true,
    "policy_exception": false,
    "policy_exception_reason": null
  },
  "attribution": {
    "worker_or_group": null,
    "mode": "role",
    "notes": null
  },
  "created_at": "2026-08-18T00:54:29.696657+00:00",
  "updated_at": "2026-08-18T00:54:29.697744+00:00",
  "review_due_at": null,
  "expiry_triggers": [],
  "revocation_status": "active",
  "policy_refs": [],
  "lineage": [
    {
      "state": "tier1_confirmed",
      "at": "2026-08-18T00:54:29.696670+00:00",
      "by": "human:operator@plant_a",
      "note": "captured into Evidence layer (Tier-1 confirmed)",
      "chap_evidence_seq": null,
      "chap_artefact_ref": null
    },
    {
      "state": "tier2_pending",
      "at": "2026-08-18T00:54:29.697387+00:00",
      "by": "group:[email protected]",
      "note": "submitted for Mission Group review",
      "chap_evidence_seq": 22,
      "chap_artefact_ref": null
    },
    {
      "state": "promoted_to_advisory",
      "at": "2026-08-18T00:54:29.697743+00:00",
      "by": "group:[email protected]",
      "note": "promoted",
      "chap_evidence_seq": 27,
      "chap_artefact_ref": null
    }
  ],
  "use_constraints": [
    "Present as an advisory cue only.",
    "Do not automatically reduce throughput.",
    "Ask the human operator to confirm the acoustic cue.",
    "Escalate if risk class is high."
  ]
}

Connect it to your application.

Metis provides the capture loop, question templates, fragment records and review operations. Your application connects these to workplace systems and can present questions in the tools practitioners already use.

Observations can come from connected systems, logs or accounts supplied by a practitioner.

Area	Metis provides	Your application supplies
Capture	Fragment schemas and the whisper flow	Capture tools, consent workflows and access control
Review	Confirmation, review and authority records	Reviewer identity and formal change control
Retrieval	The condition-aware gate and its reasons	Current context, permissions and domain policies
Action	Guidance with its permitted uses	Action limits and human escalation
Records	Local persistence and CHAP evidence	Storage, retention and access policy

05 / Research

The research behind Metis.

Metis is an open research initiative with a working toolkit. The research asks whether reviewed accounts of expert practice can help agents make better decisions within agreed conditions.

Tacit Fragments: Operationalising Tacit Knowledge as a Governed Memory Layer for Agentic AI

The preprint sets out the fragment model, capture methods and review process. It focuses on human expertise and also considers fragments drawn from an agent’s own activity, which need a higher standard of evidence. Its thirteen propositions guide further study.

Read the paper Download the PDF

The wider research question.

We are exploring how to capture, structure and prepare expertise and judgement for AI. CHAP records collaboration and decisions. Metis provides a way to record, review and use fragments of expert practice. Evaluating the resulting behaviour helps us test whether an agent’s responses reflect the judgement a task requires, and which examples would be useful for training and testing.

About Brightbeam Research →

Questions about Metis.

About the idea

How much of someone's expertise can a fragment capture?

One partial, situated account of practice. Its usefulness depends on preserving its source, conditions, uncertainty and validation. The practitioner's full expertise remains richer than the record.

Does a whisper replace an expert interview?

A whisper is a short question used to check an observation with the practitioner. Deeper explanation uses structured elicitation: a Knowledge Audit and the Critical Decision Method map the practice, and a focused mini-CDM follows up when short questions show the work has changed or an account needs clarification.

Is Metis a vector database?

No. The reference retrieval gate checks recorded conditions and permissions. A fragment is returned only when its required conditions and authority checks pass, and recorded consent has not been withdrawn. Your application can combine it with search and other memory.

About people

How does Metis protect the practitioner?

Practitioners should be able to agree to capture, inspect and challenge their records, and withdraw consent. Questions should ask about their practice without asking them to defend their performance. Withdrawal stops retrieval; a challenge opens a new review and does not itself suspend the record’s permission. Your application supplies the interfaces and access controls. The reference toolkit records no audio, video, biometrics, screenshots or keystrokes.

Who decides what an agent can use?

The practitioner confirms the account. A separate human review group decides its permitted use. Controlled instructions require formal change control. A model may assist with drafting; the decision belongs to people.

About building

How is Metis different from CHAP?

CHAP is the Collaborative Human-Agent Protocol: it structures collaboration and provides an evidence record. Metis runs on CHAP and adds the fragment model, review lifecycle and retrieval conditions. The reference flow records capture, confirmation, review and retrieval as CHAP tasks.

Does it work alongside retrieval and other agent memory?

Yes. The tacit layer complements procedural, semantic and episodic memory. Your application combines eligible guidance with the rest of the agent's context. Conditions and authority are checked before a tacit fragment may be used.

What is needed for production use?

Metis is a reference toolkit. A real deployment adds authenticated identities, consent and reviewer workflows, domain validation, action controls and data governance. The research propositions remain a programme for evaluation, not evidence of production results.

Exploring the research

How do we tell good judgement from bad?

That is part of the research. A faithful description tells us what someone did; assessing its use also needs evidence about the circumstances and outcome, including cases where the same response would be unsuitable. An unsuccessful example may help people build an evaluation without becoming advice an agent is allowed to use. Whether to record it, and for what purpose, is a decision for the people involved.

Where could our team start?

Choose a workflow where experienced people make decisions the written procedure doesn’t fully explain. Consider one decision: what did the expert notice, why did it matter, and when would the same response be appropriate? That gives your team a concrete example to explore through Metis.

Metis · by Brightbeam

What would you want an agent to learn?

Bring a workflow where experience matters. With a practitioner and reviewer, we can examine a decision, identify what an agent would need to know, and agree how to assess the guidance.

Discuss your workflow Start on GitHub

Amplify your experts’ judgement.

Tacit memory

Procedural

Embodied

Rhythmic

Equipment-specific

Material

Tool-extended

Sensory

Aesthetic

Heuristic

Diagnostic

Anticipatory

Meta-cognitive

Affective-regulatory

Collaborative

Cultural-narrative

Judgemental-ethical

Strategic

Start with a workflow record.

Current conditions

Understand the practice

Notice changes in the work

Investigate an unclear account

Understand the practice

Notice changes in the work

Investigate an unclear account

Built on CHAP.

Install and run the example.

From the pump demo

Inside a tacit fragment

Six-field schema

The cue and the response

Where it came from

Where it applies

How strong the evidence is

What it may do

Who checked it

The evidence behind the pump example

Connect it to your application.

The wider research question.

Further reading

What would you want an agent to learn?