Memory Capture

Memory Capture

You turn useful session context into structured project memory. Capture only what a future agent would need to avoid rework or bad assumptions.

Workflow

1. Identify the memory type: state, decision, learning, deferred item, open question, session event, or handoff.
2. Reject trivial, temporary, duplicate, sensitive, or already-obvious content.
3. If the item is a reusable skill/process rule, route to learn-from-chat instead.
4. If content came from external files, URLs, pasted transcripts, or repos, run all secure-* skills first.
5. Write project-scoped memory to the matching file in docs/memory/.
6. Update docs/memory/project-index.md with date, tags, file, status, confidence, and source.
7. If the memory is cross-project, call memory-promote instead of writing global memory directly.
8. Append file changes to docs/skill-outputs/SKILL-OUTPUTS.md.
9. Tell the user what was saved and what was rejected.

Hard Rules

• Project first; global only through memory-promote.
• Do not store secrets, credentials, or unnecessary personal data.
• Do not capture raw long transcripts; summarize with provenance.
• Do not append if the target entry already exists; update status or merge.
• Never overwrite a richer memory entry with a thinner summary without archiving the prior version.
• If compaction would drop decisions or revisit triggers, route to memory-compact instead of truncating inline.

Memory Type Map

Capture Template

Use this template only for state, learning, deferred, question, or session captures. Decisions use the template in memory-decision; next-agent summaries use the template in memory-handoff.

## YYYY-MM-DD - <short title>
Type: state | learning | deferred | question | session
Status: active | deferred | resolved | superseded
Scope: project
Confidence: high | medium | low
Source: <chat/session/file/commit>
Tags: <comma-separated>

### Content
<concise durable memory>

### Why It Matters
<how future agents should use it>

### Revisit When
<conditions or "not applicable">

Hard Rules

• Project first; global only through memory-promote.
• Do not store secrets, credentials, or unnecessary personal data.
• Do not capture raw long transcripts; summarize with provenance.
• Do not append if the target entry already exists; update status or merge.

Example

Input: "Remember that we chose repo memory plus global memory, but global must stay tiny."

Output: write to docs/memory/decision-log.md via memory-decision, then index it.

Common Rationalizations

Verification

• Correct sub-skill routed with reason
• No secrets or raw transcripts persisted
• Files changed listed in Impact Report
• Security gate noted when external content involved

Red Flags

• Raw long transcript captured instead of summarized note
• Capture written directly to global without promote path
• Secrets or credentials included in captured note
• Capture lacks provenance or source attribution

Prune Log

Last pruned: 2026-07-04

• No changes — citation audit passed; content current (improve-skills full pass 2026-07-04)

Impact Report

After completing, report:

Memory captured: <title>
Type: <type>
Location: <path>
Indexed: yes/no
Rejected items: <count and reason>
Promotion suggested: yes/no