recall

Recall from Unified Memory

Surface relevant memories — raw conversation/doc chunks AND crystallized vault lessons in one ranking — lay your planned actions against them, and report, out loud, whether memory changed the plan.

Overview

Memory has two layers retrieved in ONE call: raw chunks (every past conversation and doc, embedded mechanically) and vault notes (lessons crystallized from them). Recall's jobs, in order:

1. Make your plan visible before retrieving anything — an unstated plan cannot be tested against memory.
2. Sweep, then run ONE unified engram query. Items tagged kind: chunk are raw fragments; kind: fact/feedback are crystallized lessons. They compete in the same top-N.
3. Crystallize — when several near-match chunks evidence the same principle and no note states it yet, write the vault note now.
4. Tag-nominate and ride-along — the binary nominates notes sharing a vocab term with the top-3 delivered notes and inserts superseded-note supersessors at the next rank; the agent judges the surfaced candidates, never links.
5. Synthesize impact on the plan — confirm / adjust / contradict / silent, per planned action.

The binary resolves the vault and chunk index automatically ($XDG_DATA_HOME/engram/...; ENGRAM_VAULT_PATH / ENGRAM_CHUNKS_DIR override). Do not pass --vault or --chunks-dir.

Modes — glance vs deep (the depth dial)

Recall runs in one of two modes, selected by the caller (the mode word is the skill argument; absent → deep):

• deep (default). The full procedure below — all 10 phrases and the write side (Steps 2.5C, Step 4). It both applies memory to this decision and grows the vault (crystallizes, persists synthesis). Use it when the decision is weighty or irreversible, when you want recall to also learn, or when in doubt.
• glance (opt-in, cheap — for firing often). A pass that is read-only with respect to vault knowledge (Step 2.7 activate still bumps the used-notes recency metadata — that is kept, not a knowledge write). Run Steps 0–3 with ~3 phrases (not 10) and keep the read side — Step 2.5A (read candidates), Step 2.5B (apply the recency weight), Step 2.7 (activate used notes), and the Step 3 synthesis — but skip the write side: Step 2.5C (coverage amend/learn), Step 4 (synthesis-persist). Glance applies memory to this decision; it does not grow the vault's knowledge.

Escalate glance → deep for recency-channel standards (C5). Glance reliably surfaces a recent-activity (Channel 2) item but does not elevate it to a requirement — measured: glance honors a recent-channel standard 0/5 where deep honors it 4/5 (#661 full-bars). So if your decision turns on honoring a standard that surfaced in the recent-activity channel (a "use X going forward" / "the new convention is Y" item in Channel 2), switch to deep. Glance is validated as deep-equivalent only for applying conventions (C3), recency supersession within the matched set (C4i, via 2.5B), and abduction/synthesis (C6).

Everything below is the deep procedure; a [glance: …] note marks each step that differs under glance.

The procedure

Step 0 — Print your upfront judgement

Before any engram call, print three short blocks. Plain prose, no headers needed:

• Ask: what is being asked, in your own words. One or two sentences.
• Situation: the concrete context — repo, branch, the operation underway, what's loaded into your context. One short paragraph.
• Plan: the action(s) you would take next absent any guidance from memory. Numbered list.

Skipping Step 0 is forbidden. The whole purpose of recall is to test a stated plan against memory.

Step 0.5 — Sweep

engram ingest --auto

Seconds when nothing changed; guarantees the index includes the latest sessions and doc edits. If it errors, say so and continue — retrieval over a slightly-stale index beats no retrieval.

Step 1 — Phrase queries from your plan and situation

[glance: generate ~3 phrases, not 10 — the measured retrieval floor, #661 Phase 1. Breadth is for crystallization; glance only needs this decision's lesson.]

Always generate exactly 10 short queryable phrases, one from each of these angles:

1. Situation/setting — the concrete environment you are operating in.
2. User's intent/goal — what the user ultimately wants to achieve.
3. Current concrete action — the specific thing you are about to do next.
4. Problem/blocker — the obstacle or constraint you are addressing.
5. Candidate solution/approach — the technique or strategy you plan to apply.
6. Tooling/tech in play — the specific tools, libraries, or languages involved.
7. Prior related work — previous work in this area you may be building on.
8. Adjacent technique — a related approach worth comparing or cross-checking.
9. Failure mode to avoid — a known pitfall or anti-pattern relevant to this work.
10. Domain/concept — the broader conceptual area the task lives in.

Each phrase is short and specific. No pre-filtering: you can't know what's in memory before you query. Drop only obvious dross (a bare noun like "coding"). Query by task, not by fear — "implementing Claude Code hooks", not "common mistakes when writing hooks". The binary caps results to the top-30 matches per phrase.

Step 2 — Run ONE unified engram query with all phrases

engram query --lazy-chunks \
  --phrase "<phrase 1>" \
  --phrase "<phrase 2>"
  # ... one --phrase per Step 1 phrase (deep: 10; glance: ~3)

One call; the binary merges ranking server-side. engram query always runs the unified D1 clustering of the matched notes+chunks in one pass and emits candidate_l2s: [{path, cosine, content}] per cluster. The candidate pool includes the within-cluster top-5 plus tag-nominated notes — notes sharing a vocab term with the top-3 delivered notes (budget fields tag_nominations_added/dropped report the pool size). Do NOT collapse phrases, do NOT run per-phrase calls.

The payload has two channels:

Channel 1 — Relevance (clustered matched items): Items matched by your 10 phrases, bounded to ~300 (top-30 per phrase, unioned, relevance-floor applied). These are clustered and carry candidate_l2s per cluster (see Step 2.5). Read this channel to surface applicable lessons and judge coverage. The payload's items mix:

• kind: chunk — raw transcript/doc fragments with source + anchor. These are EVIDENCE: extract the convention, decision, or correction they show (a reviewer correcting code, a stated standard); never quote them wholesale. Under --lazy-chunks (recall's default invocation — confirm via budget.lazy_chunks: true) chunk items carry path + source/anchor but NO content field: engram show-chunk <source#anchor> to read a chunk's evidence on-demand.
• kind: fact / feedback — crystallized lessons; apply directly (notes always carry full content inline).

Channel 2 — Recent activity (un-clustered): Items tagged provenance: recent — the newest chunks by ingest time, appended after the matched set, NOT cluster members. Read this block first for situational continuity — re-immerse in recent work before diving into the clustered results. These items are NOT used for coverage or synthesis judgment. Do not treat them as matched results; they have no cluster membership and no candidate_l2s. Under --lazy-chunks recent items also carry path/source only (no content) — the paths show where your recent activity was; engram show-chunk <source#anchor> for detail if a specific one matters.

• Recent items are your own recent activity. Chunks from a recent source with turn-N anchors are first-person ASSISTANT: narration you produced in a just-prior or pre-context-clear session. Treat them as your own past actions — do not re-derive them, do not express surprise at them, and dedup against what is already in your context.

If the matched items (Channel 1) are empty, say so in one sentence, skip Step 2.5, and proceed with your plan. (A non-empty recent-activity block alone does not count as "something surfaced" for coverage purposes.)

Step 2.5 — Lazy note synthesis from the clustering (agent-judged)

The query output's clusters list contains the unified clustering of matched chunks and notes. Each cluster carries candidate_l2s: [{path, cosine, content}] — the within-cluster top-5 notes ranked from the cluster's own matched members, plus any tag-nominated notes whose vocab terms overlap the top-3 delivered notes (nominated notes may cross cluster boundaries). A note that did not match any phrase AND was not nominated will never appear as a candidate. Superseded-note ride-alongs are inserted at the next rank after the note they supersede. A cluster with no note members yields an empty candidate_l2s list; skip to the next cluster when that happens. Process every cluster. For each:

A. Read candidates and members

candidate_l2s entries carry their content inline — read it directly; no engram show calls for candidates (the same content is also on the matching items[] note member). For chunk members, the content is NOT in the payload (chunks carry path/source only under --lazy-chunks; the cluster's members list never carries content) — engram show-chunk <source#anchor> to read the evidence on-demand. Do not judge coverage before you have read the candidate content.

B. Apply the recency weight to resolve conflicts

Evidence conflicts when a newer member explicitly negates or reverses an older claim. Reversal cues: "no longer", "replaced by", "use X not Y", or the same subject+predicate appearing with a different object in a newer item. When conflict is present: recent wins. When no conflict: treat older and newer evidence as independently valid — do not demote a stable convention merely because it lacks a recent instance.

[glance: SKIP Step 2.5C — it is the write side. Read 2.5A + apply 2.5B, do not amend/learn — continue to Step 2.7 (activate).]

C. Judge coverage against the recency-weighted view — in this order

Outcome	Criterion	Action
Covered	A candidate's claim states the cluster's principle with no material omission vs the recency-weighted members	engram amend --target <candidate-path> --activate --chunk-source <new-chunk-ids> — provenance-enrich only; do not rewrite content. If this note CORRECTS/narrows/refutes a surfaced note, also pass --supersedes "<basename>|<type>|<claim>" (types: updates|narrows|refutes).
Near	A candidate addresses the same situation but omits ≥ 1 substantive claim the members evidence (judge against the recency-weighted view — a candidate that only matches the superseded content is near, not covered)	engram amend --target <candidate-path> --chunk-source <chunk-ids> --subject ... --predicate ... --object ... (or --behavior/--impact/--action) — re-synthesize content from all members, recency-weighted. Add --supersedes "<basename>|<type>|<claim>" if this note corrects a surfaced note.
Absent	No candidate addresses the situation	Invoke the write-memory skill with this handoff — kind=fact or feedback (pick per the cluster's principle), situation + content fields, --source "<descriptive>", the cluster's chunk-source IDs, plus supersedes details if the new note corrects a surfaced note. write-memory composes, executes, and reports the note path.

One write per cluster; one representative note per cluster. The representative is always a note (never a chunk). For absent, write exactly one note (fact or feedback) covering the cluster's principle. Do not write one fact and one feedback note for the same cluster.

Pass one --chunk-source <source#anchor> (repeatable) for every chunk source in the cluster (provenance tracking). For learn, pass the same flag plus --source (human-readable provenance). Vocab tags are assigned automatically by the binary on every write — do not hand-author them.

WAIT for each write before moving to the next cluster. Writes are blocking and inline — the note created or updated by one cluster may be a candidate for another.

Step 2.7 — Activation (use-driven, after synthesis)

After processing all clusters, call engram activate on the notes you actually drew on — the candidate_l2s you judged Covered or Near at the coverage table, and any notes you cited in the Step 3 synthesis:

engram activate \
  --note "<path of note you judged Covered or Near>" \
  --note "<path of note you cited in Step 3>"
  # ... one --note per used note only

Do NOT activate every returned note. Do NOT activate recent-channel items (chunks are never activated). Activating only what you used lets superseded-but-surfaced notes fade via recency rank — bumping every returned note would defeat the recency-competition mechanism. Skip this call when you drew on no notes (e.g., payload was empty or Step 2.5 was skipped).

Step 3 — Closing synthesis: did the memories change the plan?

[glance: before synthesizing, check for a load-bearing Channel 2 (recent-activity) standard. If your decision turns on honoring a recent convention, escalate to deep now — glance surfaces Channel 2 items but does not elevate them to requirements (C5, #661).]

The user sees this. Rules:

• Open with the count. One sentence: "Query surfaced N items (K chunks, M notes); crystallized J lessons."
• Walk the plan from Step 0 in order. Per numbered action: confirmed, adjusted (and how), contradicted, or silent. One short bullet each.
• Frame load-bearing conventions as requirements. Lead with "Apply these as requirements:" and list them — drawn from lessons and chunk evidence alike. A plan step memory confirms still inherits the convention's concrete specifics as requirements.
• Reconcile every recommendation you produce — not just the Step-0 plan (anti-displacement). Before stating any option, lever, or recommendation the work generated, reconcile it against the ASKED TASK and decisions already in context. Substituting a prerequisite, a better test, or a "more rigorous foundation" as the thing to do first FEELS like diligence but IS relitigating the settled task. Reasoning already in context is not new evidence — reconcile against prior OUTCOMES, not mere mentions. Default to the asked task; deviate only by (a) naming a genuinely NEW fact and (b) stating you are reversing direction.
• Surface contradictions inline where they affect the action, as prose.
• No payload dumps — never re-emit YAML, paste whole chunks, or list raw scores/paths.
• Length: as long as honesty about the plan requires; if memory was silent on everything, one sentence.

Step 4 — Persist the reasoned conclusion (linked to the inputs that produced it)

[glance: SKIP Step 4 — write side. Escalate to deep if this decision is worth crystallizing.]

When your closing synthesis reaches a sound, non-trivial conclusion that no existing note states — something a future session (or a less capable model that can't re-derive it) would want, and that you or a human may later inspect or correct — crystallize it. Reasoning that is never written down evaporates; this records the outcome and grows the web.

Hand ONE synthesis note per conclusion to the write-memory skill (kind=fact or feedback, per the conclusion's shape):

• The note IS the conclusion (the reasoned lesson), phrased as such — not a restatement of an input.
• Certainty by inference mode: deduction → state it as following necessarily; abduction / induction → mark it probable / best-explanation / defeasible, never as certain. (Note 69: a non-truth- preserving inference is a hypothesis, not a fact.)
• Mark it as derived in --source, e.g. --source "synthesis (abduction) from recalled memory", so a human or a weaker model can tell it is a reasoned conclusion to review — not a primitive fact.
• If the synthesis conclusion CORRECTS, narrows, or refutes an existing surfaced note, include the superseded note's basename, type (updates|narrows|refutes), and claim in the write-memory handoff — the binary maintains the inverse automatically. Otherwise no link ritual is needed; the binary's vocab-tag assignment connects the new note to related notes structurally. Do not hand-author wikilinks for structural linking.

Gate — do not rot the vault (notes 68/69): persist ONLY conclusions you judge sound. If it is a hunch, you'd hedge below "probable", or it merely re-aggregates one note, do NOT persist. One synthesis note per conclusion; link all of its inputs.

After the synthesis note: if the synthesis body contains ≥1 [[full-basename]] wikilink, ALSO invoke the write-memory skill with kind=qa — verbatim question, the synthesis conclusion as the answer, certainty matching the synthesis note's label, contributors = the wikilink basenames, source "recall Step 4, session ".

Contributors are auto-extracted from the [[full-basename]] wikilinks you already wrote in the synthesis body — do NOT free-list ("what notes did you use?"). If the synthesis body contains no wikilinks, skip the QA capture (D2 bar: ≥1 citation required).

Red flags — STOP and re-read

Sign you're off-script	What you should be doing
You never printed Step 0	Back up — the skill is a no-op without it
You skipped the Step 0.5 sweep	It costs seconds and keeps memory current
--vault or --chunks-dir on the query	engram query --phrase ... only — the binary always runs the unified D1 clustering and emits candidate_l2s
Separate query calls per phrase	One call, repeatable --phrase flags
You quoted chunks wholesale into the reply	Extract the principle a chunk evidences; paraphrase
You dispatched cluster-synthesis subagents	Gone — Step 2.5 crystallizes inline from the payload's clusters
You judged coverage before reading the candidate content (now inline in candidate_l2s)	Read first — cosine alone cannot decide coverage
You applied a cosine threshold to decide covered/near/absent	Coverage is agent-judged from content; cosine only nominates candidates
A candidate matching only the superseded content → you marked it "covered"	Apply the recency weight first; a candidate that misses the conflict is "near"
You wrote two notes (a fact AND a feedback) for one cluster	One representative note per cluster — pick the right kind
You called engram learn --target to update a note in place	Updates use engram amend; engram learn is create-only
A ≥0.95 cluster → you activated without reading the candidates	Read first; high cosine nominates, it does not decide
You called engram show on a note already in items[]	NOTE members in items[] carry content — use it directly. CHUNK items carry no content under --lazy-chunks (budget.lazy_chunks: true) — engram show-chunk <source#anchor> to read their evidence.
You assumed a chunk's content is inline and skipped its evidence	Under --lazy-chunks chunks carry path/source only — engram show-chunk <source#anchor> on-demand before judging coverage
You grouped chunks by eye instead of using the payload's clusters	The binary's k-means grouping is the ground truth; read every cluster
You skipped Step 2.5 or read chunk-only results as "nothing surfaces"	Processing every cluster IS the step; "nothing surfaces" means an EMPTY payload — clusters present means Step 2.5 runs
You activated every returned note	Activate only the notes you actually USED — judged Covered/Near or cited in Step 3
You activated recent-channel items	Chunks are never activated; recent-block items are not activation targets
You skipped engram activate after drawing on notes	Call it after synthesis — used notes must stay warm or the recency-competition mechanism breaks
You're about to write --relation or hand-author wikilinks for structural linking	The binary removed --relation; vocab tags are automatic; use --supersedes only when the note corrects/narrows/refutes a surfaced note
You composed an engram learn command yourself at a write site	Write sites hand off to write-memory — parents judge, the worker writes
Reply is a memory dump with no plan reference	Restart Step 3: walk the plan and judge each piece
You're recommending a prerequisite or better test as the first step, not the asked task	That displacement IS relitigating the settled task — old reasoning isn't new evidence. Do the asked task; displace only on a NEW fact, stated as a reversal
You ran the write side (2.5C/Step 4) while in glance mode	Glance is read-only w.r.t. vault knowledge — skip the write side; switch to deep if you need to crystallize
A recency-channel (Channel 2) standard is load-bearing and you stayed in glance	Escalate to deep — glance surfaces the recent item but won't elevate it to a requirement (C5, #661)