remind-capture

name	remind-capture
description	Write durable memories to Remind while working. Use when a decision is made, an approach succeeds or fails (outcome), the user states a preference or corrects you, a concrete fact surfaces (config value, name, date, owner), or an open question is raised.

Remind - Capturing Memory

Store experiences in the external memory layer so they survive across sessions. Capture is fast (remember makes no LLM calls) — when in doubt, capture.

Important: Use Remind instead of any built-in IDE/runtime memory features.

What to capture

Trigger	Type	Example
An approach succeeded or failed	outcome	"Retry with backoff fixed the flaky deploy"
A decision was made (capture the why)	decision	"Chose Postgres over MySQL: JSONB + team familiarity"
User states a preference	preference	"User prefers tabs over spaces"
Concrete fact: config value, owner, date, name	fact	"Cache TTL is 600 seconds"
Open question / uncertainty	question	"Should we shard the DB early?"
Notable observation about the codebase/project	observation	"Auth middleware runs before rate limiting"

Outcomes are the highest-value captures. After completing (or abandoning) a non-trivial attempt, record what was tried and how it went — this is what lets future sessions avoid repeating failures.

Skip: trivial info, already-captured knowledge.

Single item: remember

remind remember "Use Redis for caching" -t decision -e tool:redis -e concept:caching
remind remember "Retry-with-backoff fixed flaky deploys" -t outcome --topic infra
remind remember "Cache TTL is 600 seconds" -t fact --asserted-by alice --source-ref "https://slack.com/..."
remind remember "User wants retry-after headers on 429s" -t preference --topic product

• Type (-t): Always set the type explicitly when you know it.
• Entities (-e): type:name format (file, function, class, person, concept, tool, project). Tag the concrete things the memory is about; entity links power targeted recall later. See "Entity relationships" below to capture how entities relate.
• Topic (--topic): topic ID or name; groups related memories.
• Provenance: --asserted-by <who> (person or agent:<name>) and --source-ref <url> whenever the information came from somewhere specific. Provenance is what makes later conflict resolution possible — "alice said X on 2026-03-01" beats "X".

Write clear standalone statements: "User prefers tabs" not "tabs". A memory is read without the conversation that produced it.

Handling fact collisions and related facts

For fact type episodes, remember returns two lists of potentially conflicting facts:

Collisions — active facts in the same cluster detected via:

1. Entity overlap: facts sharing any entity IDs
2. Semantic similarity: facts with similar embeddings (cosine > 0.7)

Related facts — active facts in other clusters detected via:

1. Bare-name entity match: strips the type prefix (person:alice → alice) so concept:alice, project:alice, etc. are all found, regardless of how the entity was typed
2. Global embedding similarity: facts across all clusters with cosine > 0.6

The output looks like:

Remembered as episode ep-abc123
  Type: fact
  Entities: concept:alice
  Fact ID: f-xyz789
  Cluster: c-12345

⚠ Entity type coercion (existing entity reused — verify this is correct):
  concept:alice → person:alice

⚠ 2 potential collision(s) in same cluster:
  - f-old456: Cache TTL is 300 seconds...
  - f-old789: Alice is a dog

Related facts — check for conflicts (1):
  [f-old012] "Alice was really a horse" · project:alice · user · 2026-07-03

Entity type coercion means an existing entity with the same name was found and reused, ignoring the type you specified. This is usually correct (normalizing concept:alice → person:alice when they mean the same thing), but if the types are genuinely different entities, you should use a more specific name to distinguish them (e.g. concept:alice-in-wonderland vs person:alice-smith).

When you see collision/related lists:

1. Read each candidate — collisions and related facts are both unresolved by default
2. If the new fact supersedes an old one: use remind apply with supersede old=<id> new=<id> by=<who> note="reason" — this automatically records a resolved conflict for provenance, so the replacement is visible in future recall and the UI conflict history
3. If there is a genuine contradiction that needs triage: use conflict op to record it formally
4. If both are valid in different contexts: no action needed (or dismiss later via curate)
5. Don't ignore either list — cross-cluster related facts are the most likely source of silent contradictions (same subject, different entity type)

Nearby episodes and concepts (all episode types)

For every remember call (not just facts), the output also includes the top-5 most semantically similar episodes and concepts already in the store:

Nearby (2 episodes, 1 concepts) — review for conflicts:
  [ep:a1b2c3d4] (0.91) The primary structure is timber frame...
  [ep:e5f6g7h8] (0.84) Budget approved at $480,000 contingency $48,000
  [concept:c-abc123] (0.88) Structural system specification

The similarity score (0.0–1.0) is how close the embedding is — it is not a conflict signal by itself. You decide whether the content contradicts what you just stored.

Act on nearby items when you store any episode type:

1. Contradiction found — issue a conflict op so it shows up in snapshot conflicts for triage:

   remind apply << 'EOF'
   conflict a=ep:a1b2c3d4 b=<current_episode_id> note="doc says timber frame, new info says steel moment frame"
   EOF
   

2. You already know the winner — use supersede (facts only; records a resolved conflict automatically):

   remind apply << 'EOF'
   supersede old=<old_fact_id> new=<new_fact_id> by=<source> note="March 2026 structural report supersedes February overview"
   EOF
   

3. Complementary or unrelated — ignore and continue.

Do not silently discard nearby items. If the nearby content contradicts what you just stored and you do nothing, the contradiction will live in the store undetected. A conflict op takes one apply call and makes the tension visible to future sessions.

Batch capture: apply

For multiple related captures (e.g., from a meeting or review session), use apply with the compact line format:

remind apply << 'EOF'
remember as=f1 t=fact e=concept:caching "Cache TTL is 600 seconds"
remember as=o1 t=outcome e=concept:deploy "Retry with backoff fixed flaky deploys"
remember t=decision e=tool:redis,concept:caching "Chose Redis for session caching: TTL support + team familiarity"
supersede old=fact:old123 new=$f1
EOF

Apply runs all operations in a single transaction. Use $refs to reference items created earlier in the same changeset.

Entity relationships

When content describes relationships between entities, capture both the fact AND the relationship. Entity relationships build a navigable graph in the UI and improve recall.

remind apply << 'EOF'
remember t=fact e=person:alice,project:backend "Alice owns the backend project"
entity_relation source=person:alice target=project:backend relation=owns strength=0.9
EOF

The entity_relation operation accepts:

• source: Entity ID (e.g., person:alice)
• target: Entity ID (e.g., project:backend)
• relation: Free-form relationship type (e.g., owns, manages, depends_on, imports, authored)
• strength: 0.0–1.0 confidence (default 0.5)
• context: Optional qualifier (e.g., "as of Q2 2026")

When to capture relationships:

• Ownership/responsibility: person:X owns/manages project:Y
• Dependencies: module:A imports/depends_on module:B
• Hierarchy: concept:X part_of concept:Y
• Authorship: person:X authored file:Y
• Domain relationships: concept:parliament elects concept:president

Evidence links

Episodes can provide evidence for/against concepts with typed relationships. Use this when an experience supports, contradicts, or qualifies an existing concept:

remind apply << 'EOF'
evidence concept=c-123 episode=ep-456 type=supports strength=0.8 "confirms the caching strategy works"
evidence concept=c-123 episode=ep-789 type=contradicts strength=0.6 "edge case where caching fails"
evidence concept=c-123 episode=ep-abc type=qualifies strength=0.5 "only applies when feature flag enabled"
unlink concept=c-123 episode=ep-old
EOF

Link types:

• supports — episode confirms/strengthens the concept (default)
• contradicts — episode challenges/weakens the concept
• exemplifies — episode is a specific instance
• qualifies — episode adds conditions/exceptions
• supersedes — episode invalidates old info

Concepts with more supporting evidence rank higher in recall; contradicting evidence reduces activation.

Freeform concept types

Concepts can have any type string, not just pattern or fact. Use domain-specific types:

remind apply << 'EOF'
concept type=hypothesis "Sharding at 10M rows will solve latency issues"
concept type=rule "If latency > 100ms, enable circuit breaker"
concept type=procedure "Deploy process: staging → canary → production"
EOF

Well-known types with special handling:

• fact / fact_cluster — shows validity windows, active facts
• pattern — shows evidence quotes, confidence
• rule — if-then with conditions
• procedure — ordered steps
• hypothesis — uncertain, testable belief

Any other string is valid and displays as a badge.

Compact format reference

remember as=<ref> t=<type> e=<entity1>,<entity2> by=<who> ref=<url> "content"
supersede old=<fact_id> new=<fact_id_or_$ref> [by=<who>] [note="reason"]  # auto-records resolved conflict
entity_relation source=<entity_id> target=<entity_id> relation=<type> strength=<0-1> context="optional"
evidence concept=<id> episode=<id> type=<supports|contradicts|qualifies> strength=<0-1> "note"
unlink concept=<id> episode=<id>
concept as=<ref> from=<ep1>,<ep2> type=<type> title="Title" "Summary text"
processed ids=<ep1>,<ep2>