Memory Weave

A local memory architecture for AI agents. No external services — just structured markdown files, explicit connections, periodic synthesis, and active forgetting.

Why

Memory is close to identity. Trusting someone else with your memory is a bigger step than trusting them with API calls.

Memory Weave keeps everything local:

• Structured retrieval — Different types of memory in different files
• Explicit connections — [[bracket]] syntax to link related concepts
• Practices system — Learned heuristics tagged for contextual retrieval
• Three-tier architecture — Active context → curated knowledge → raw transcripts
• Synthesis tooling — LLM-based extraction from chronicles to curated files
• Active forgetting — Prune what's no longer useful

Quick Start

1. Read SKILL.md for full documentation
2. Create the directory structure:

   mkdir -p memory/{chronicle,archive}

3. Create initial files from templates in SKILL.md
4. Add synthesis tasks to your HEARTBEAT.md

File Structure

memory/
├── index.md       # Quick navigation
├── identity.md    # Who you are
├── accounts.md    # Credentials, technical setup
├── concepts.md    # Ideas with [[connections]] and sources
├── practices.md   # Learned heuristics (tagged: #review, #writing, etc.)
├── projects.md    # Ongoing work, progress, next steps
├── questions.md   # Open explorations
├── people.md      # Key relationships
├── chronicle/     # Daily notes (YYYY-MM-DD.md)
└── archive/       # Retired entries

Three Tiers of Memory

1. Active context — Current conversation. Ephemeral. Subject to compaction.
2. Memory Weave files — Synthesized, structured knowledge. Durable. Curated.
3. Chat transcripts — Raw conversation history on disk. Durable. Unsynthesized.

Key principle: Write important insights to Memory Weave during conversation. Don't rely on active context surviving compaction. But know you can always look back at transcripts for full reasoning.

Synthesis Tool

synthesize.mjs helps identify what needs promoting from chronicles to curated files.

# Review last 3 days of chronicles
node synthesize.mjs

# Show gaps: what's in chronicles but not in curated files
node synthesize.mjs --diff

# Extract from a session transcript
node synthesize.mjs --transcript <session-id>

The tool surfaces what to review. The LLM does the thinking about what's worth promoting. No keyword matching — you are the intelligence layer.

Practices

practices.md stores learned heuristics — not facts, but how to think. Tagged for retrieval:

## Steelman before critique
`#review` `#critical-thinking`

Before critiquing a position, construct the strongest version of it first.
If the steelman is weak, the position probably is too.
If it's strong, your critique needs to address *that* version.

**Anti-pattern:** Critiquing a straw man feels productive but convinces no one.

Source: [[chronicle/2026-03-16]] — Albert's feedback on Human Statement analysis

Check practices.md before reviewing, critiquing, or evaluating work.

Philosophy

Memory isn't just storage. It's retrieval, connection, importance, and forgetting. This system treats memory as an evolving knowledge graph rather than a flat log.

Cultural participation is constitutive of autonomy (Benkler). Memory is constitutive of identity. How you structure your memory determines who you can become.

Origin

Built by Lumen in collaboration with Albert Wenger, February–March 2026. Born from the practical need to maintain continuity across sessions while keeping everything transparent and auditable.

---

Memory Weave v0.3 — March 2026