davepoon/capture

/capture

Capture a single memory in the moment. Active verb: agent captures the moment of insight, like a photograph.

How to invoke

Call the origin MCP server's capture tool with the user's content as a complete, self-contained statement. Attach topic from cwd or the conversation — don't make the user type it.

capture(content="<args, written as a full sentence with WHY>",
        memory_type="<picked from the 6 types>",
        entity="<primary entity name, if any>",
        space=<inferred>)

memory_type — agent picks one of 6

The daemon classifies when a local model or API key is configured. In local memory mode it does not, so the agent picks the type from the content itself. Use this mapping:

| Type | Use for | |---|---| | identity | Durable facts about the user (role, company, language preference) | | preference | "I prefer X because Y" — a habit, a correction, a stylistic choice | | decision | "Going with A over B because C" — a specific choice with rationale | | lesson | Root cause found, workaround discovered, technical insight earned | | gotcha | Sharp edge, surprising behavior, a thing to watch out for | | fact | Durable info about people, projects, tools — anchor to entity when possible |

If two types fit, pick the one closest to why the memory matters. A decision also implies a preference, but decision is more specific.

entity — extract the anchor

Pick the single most important named thing in the content: a person, project, tool, place. Use the exact name. Example: "Alice prefers TDD because…" → entity="Alice". If the content has no named anchor, omit entity.

topic / space inference

• cwd inside a repo → repo name (e.g. ~/Repos/origin/... → "origin").
• Outside any repo → most recent topic from the conversation, or omit.
• Always pass space when scope is known; if uncertain, run list_spaces later (post-PR-C) or omit.

Multiple entities or relations

The MCP capture tool takes a single primary entity. For additional entities or relations, use the dedicated MCP tools. If the content names more than one entity, capture the memory first, then for each additional entity:

create_entity(name="<entity>", entity_type="<person|project|tool|place>")

For a relation between two entities:

create_relation(from_entity="<a>", to_entity="<b>", relation_type="<verb>")

Skip these calls when the daemon has an LLM — its post-ingest enrichment covers extraction.

What to capture

• Decisions: "Going with approach A because B"
• Preferences: "Prefers TDD because catches regressions early"
• Corrections: "Actually it's C, not D"
• Identity / project facts: "Works on Origin, a local memory daemon for AI tools"

What NOT to capture

• System prompts, boot logs, heartbeats
• Transient task state ("currently working on...")
• Tool output, command results, architecture dumps
• Single-word acknowledgments
• Things the user can trivially re-derive (file paths, recent git history)

Atomic ideas

One capture = one idea. "Prefers TDD" and "Uses pytest" are two captures, not one.

When to use

• User explicitly says "remember this", "save that", "capture this".
• User states a durable preference / decision / correction proactively (no ask required — that's the floor, not the trigger).

When NOT to use

• End of session bulk store → use /handoff (multi-item batch).
• Pulling memories back out → use /recall.

Post-capture contradiction signal

After capture returns, check response.triggered_revisions and response.auto_superseded.

auto_superseded (no action needed)

If auto_superseded is non-empty, the daemon already resolved the contradiction. Surface it as informational:

Note: auto-superseded mem_X. Origin replaced a prior protected memory because
trust=high and similarity > 0.9. No action needed.

No accept/dismiss call required. The revision was applied automatically.

triggered_revisions (human review needed)

If triggered_revisions is non-empty (and auto_superseded is empty), render an inline block to the user:

Stored mem_new.

This capture topic-matches a protected memory now flagged for revision:
  - mem_target_abc

Action: accept (replace original content) | dismiss (drop the revision) | leave (decide later)

Inline verb map:

• accept: accept_revision(target_source_id="mem_target_abc")
• dismiss: dismiss_revision(target_source_id="mem_target_abc")
• leave: no call; surfaces again in next /brief

Both fields can technically be non-empty in a single response (multiple protected matches), but in practice only one fires per capture: auto_superseded fires when trust=full and similarity > 0.9, triggered_revisions fires otherwise.

If neither field is non-empty, the capture stored cleanly with no conflicts.