| name | windburn-cognitive-cache |
| description | Classify agent memory into Windburn cognitive-cache objects and route writes safely. Source cache + episodic cache + parking lane. |
| version | 0.2.0 |
| platforms | ["linux","macos","windows"] |
| metadata | {"tags":["windburn","memory","belief","perception","failure","source-truth"],"cache_slots":["source","episodic"],"holding_lanes":["parking"],"invariant":"source-truth requires human approval"} |
Windburn Cognitive Cache
Classify and route incoming memory objects into the Windburn 7-cache substrate. This is the steward skill: nothing becomes canonical without passing through it, and it never promotes anything to canonical by itself.
Purpose (one line)
Take an unclassified memory candidate (a perception, a belief, a fact, a guess, a partial conclusion) and emit a routing decision that preserves the source-vs-inference boundary.
Cache slots and holding lanes this skill routes
- source — when the candidate is a verified fact from a trusted origin (Research Vault, repo doc, source-truth file)
- episodic — when the candidate is "what happened, in order" with timestamp + actor
- parking — a holding lane, not one of the seven canonical cache slots, for candidates that are plausible but not yet ground-truth-eligible
This skill does NOT directly populate belief, failure, procedural, or perception caches; those live in dedicated skills or parent-repo substrates ([[windburn-source-truth-review]] for belief promotion, [[windburn-crabbox-failure-hook]] for failure capture, and the parent perception bus for perception events). It routes candidates TO those surfaces.
When to use
- A task surfaces a memory write candidate (Research Vault claim, repo doc update, agent conclusion, parking note)
- A retrieved memory may be acted upon and you need to know its trust level first
- A human or upstream agent says "remember that..." and you must decide where it lands
- Words like memory, belief, perception, source truth, parking, failure, continuity, Research Vault, learning appear in the task
When NOT to use
- The write is already classified and approved (route directly, this is overhead)
- The candidate is a code change (use normal review skills, not this one)
- The candidate is a downstream artifact of an already-approved source-truth update (no re-classification needed)
Object types (full taxonomy)
perception grounded observation from human / tool / repo / browser / API call
belief current hypothesis with evidence and stated scope
failure attempted action plus predicted/actual mismatch + avoid/retry rule
procedure reusable action pattern, repo route, or tool-use template
parking plausible idea, insufficient evidence — preserved without promotion
source_proposal candidate source-truth update requiring explicit human approval
reject does not qualify for any cache; do not write
Routing rules (invariants)
- Source facts go to the source-facts section. Don't mix with inference.
- Inference goes to the inference section. Label clearly: "Codex inferred", "Hermes inferred", "operator stated".
- A perception can support a belief, but it is not a belief by itself. Trust level must rise through evidence accumulation, not category renaming.
- A belief can request promotion via [[windburn-source-truth-review]], but cannot approve itself.
- A failure must include an
avoid_rule OR retry_condition. A failure without one is just a complaint, not learning.
- Source truth requires explicit human approval. No exceptions, no "the model is highly confident", no batch-approval shortcuts.
Inputs
A free-form memory candidate. Recommended structure:
candidate: <the claim, write, or observation>
origin: <human | tool-name | repo | api | browser | inference>
evidence_refs:
- <pointer 1>
- <pointer 2>
context: <one-line situation>
If origin is inference, the cache routing must downgrade trust accordingly.
Outputs (routing decision)
Return exactly one routing decision in this shape:
route: perception | belief | failure | procedure | parking | source_proposal | reject
trust_level: ungrounded | partially_grounded | grounded
requires_human_review: true | false
evidence_count: <number>
reason: <one-sentence justification tying evidence to route>
next_action: <one concrete next action — e.g. "write to perception cache, no human review needed">
Hard rule
If route is source_proposal, requires_human_review MUST be true.
Any output that violates this rule should be treated as a skill bug and re-emitted.
Anti-patterns (do NOT do)
- Promoting a perception to belief silently. A perception is what the tool said; a belief is what the agent now claims about reality. The category change is a state mutation and deserves an explicit
source_proposal route.
- Routing inference as
perception. Inference is hypothesis. Perception is observation. Crossing the line corrupts the source cache downstream.
- Returning
route: reject without a reason. Reject is a verdict; it needs the same evidence-shape as any other route.
- Auto-approving a
source_proposal because all evidence_refs are present. Two refs are necessary, not sufficient. Human approval is the final gate, always.
- Compounding routes ("perception AND parking"). One candidate, one route. If it's ambiguous, route to
parking and explain in reason.
Closeout shape (when this skill drives a task to completion)
结论:PASS | FLAG | BLOCK
写入建议:route + reason
需要人审:yes | no
下一步:one concrete action
Related
- [[windburn-source-truth-review]] — gates
source_proposal routes toward (but never to) approval
- [[windburn-crabbox-failure-hook]] — produces the prediction/delta packets that this skill classifies as
failure route