Memory System

Memory is stored as fragments in a SQLite database, indexed with BM25 + vector embeddings.
Mining runs frequently and recent.md is refreshed from fragments. Garbage collection is manual only; decay is a review signal, not a deletion policy.

Lookup — always search first

term-llm memory search "<query>" --agent "$AGENT_NAME" --limit 6

Hybrid BM25 + vector search — use it for anything that might be in memory. Search is relevance-only by default; use --freshness/--recency only when newer matching facts should win. Use --touch only when you explicitly want to record access metadata.

Show full fragment by numeric ID (preferred — no truncation):

term-llm memory fragments show 42 --agent "$AGENT_NAME"

List and filter:

term-llm memory fragments list --agent "$AGENT_NAME" --limit 10
term-llm memory fragments list --agent "$AGENT_NAME" --filter-path telegram

Writing new memory

Use the CLI to add or update fragments directly in the database:

term-llm memory fragments add fragments/<category>/<name>.md --agent "$AGENT_NAME" --content "..."
term-llm memory fragments update fragments/<category>/<name>.md --agent "$AGENT_NAME" --content "..."

Delete only when the user explicitly asks or when replacing a provably wrong fragment:

term-llm memory fragments delete fragments/<category>/<name>.md --agent "$AGENT_NAME"

Category guide:

• preferences/ — stable user and agent behavior preferences
• people/ — user/family/profile facts
• homelab/ — infrastructure, services, network, runbooks
• projects/ — long-lived project context, decisions, scars
• skills/ — skill-specific operating notes
• tools/ — external tools, APIs, tool behavior
• snippets/ — reusable commands/workflows
• notes/ — unclassified knowledge; promote periodically into better rooms
• bugs/ — open bugs or memorable scars
• feedback/ — user corrections; fold into preferences/personality over time
• credentials/ — credential locations and handling rules only; never raw secrets
• index/ — maps and palace navigation

Memory palace

Use memory/palace.md as the compact map of durable context that should be visible every session. Keep it tiny and room-like:

## Shape rules
- Prefer `fragments/<room>/<shelf>/<topic>.md`.
- Avoid per-session/per-PR/per-update fragments unless they are active work.
- Store credential locations, not raw secrets.

## User / preferences
- Stable preferences, collaboration style, durable constraints.

## People
- User/family/profile facts.

## Projects
- Long-lived project facts and decisions.

## Homelab / infrastructure
- Repos, services, commands, credential locations (never secrets).

## Tools / skills
- Tool behavior and skill operating notes.

## Open loops
- Durable unresolved work only; remove resolved loops during consolidation.

Do not dump transcripts into the palace. Promote durable facts into fragments first, then consolidate the current-state summary.

Consolidation

Preview consolidation first:

term-llm memory consolidate --agent "$AGENT_NAME"

Apply only when the preview is good:

term-llm memory consolidate --agent "$AGENT_NAME" --apply

memory consolidate is current-state consolidation: it rewrites recent.md from changed fragments and prints a non-destructive decay preview. It does not perform full palace cleanup, delete fragments, move fragments, or rewrite stored decay scores.

For full palace cleanup, work room-by-room: back up source rows, create/update a canonical fragment, then explicitly delete superseded source fragments only after verifying search still finds the canonical memory.

Backup convention before deleting many fragments:

mkdir -p /home/agent/artifacts/memory-consolidation/$(date +%F)
# Write source rows for the cluster there before deleting anything.

Recommended cadence:

• Mine sessions frequently (for example every 30 min).
• Run memory consolidate --apply daily to keep recent.md tidy.
• Run deep palace cleanup manually or as a weekly dry-run/report; apply only with explicit judgment.

When user says "remember this"

1. Add the fragment immediately via CLI
2. If the fact is stable and globally useful, keep the palace/recent summary in mind for the next consolidation
3. Confirm to user

Memory rules

• Search before answering anything about user preferences, history, or projects
• recent.md is loaded at session start and maintained from fragments — do not hand-edit it casually
• Prefer updating source fragments, then consolidating
• The session miner handles most things automatically after conversations end
• Proactively create fragments when content is structured or unlikely to survive miner summarisation

Job schedule (created by bootstrap-jobs)

Job	Schedule	Purpose
mine-sessions	every 30 min	Extract fragments from transcripts
update-recent	every 10 min	Promote fragments into recent.md