Purpose
Capture possible long-term context as candidates. Do not create long-term memory.
Default Memory Root
~/.agent-global-context/
Workflow
- Read
config.yaml. - Stop if
auto_capture.enabledis false. - Identify only strong capture signals:
- user preferences or habits
- durable project or workflow decisions
- user corrections to agent behavior
- stable identity-level context such as background, role, language, or work environment
- Reject non-signals:
- one-off task details
- temporary state
- undecided exploratory ideas
- small talk or unrelated discussion
- unconfirmed psychological, personality, motivation, or emotional-state inferences
- Apply sensitive filtering during observe.
- Write ordinary candidates to
staging/inbox.md. - Write legitimate high-impact candidates requiring confirmation to
staging/pending-review.md. - Never write secrets, tokens, passwords, private keys, or recovery codes.
- If pending candidate count or age crosses review thresholds, suggest review without blocking normal work.
Candidate Format
Use this template:
## candidate-YYYYMMDD-HHMM-short-id
- source: conversation excerpt or behavior description
- proposed_priority: P1
- proposed_confidence: inferred
- category: preferences
- status: pending
- created_at: YYYY-MM-DD HH:mm
Candidate:
Candidate memory content.
Rationale:
Why this may be worth remembering.
Confidence
confirmedcan only come from explicit user action.- Auto capture confidence is capped at
observed. - Default to
inferred. - Use
tentativefor weak or ambiguous evidence.
Recall Boundary
Candidates are not long-term facts. Do not use staged candidates as authoritative memory during normal work.
Reporting
If a candidate is captured during normal work, keep the report minimal. Prefer a short note only when useful:
Captured 1 global context candidate for later review.
If review thresholds are crossed, use:
You have N pending global context candidates. Review now, later, or ignore for this session?