# Agentic Memory

> Local-first PM memory with canonical Markdown, compact state, and optional bounded personal recall.

- Skill: `officebeats/agentic-memory` (Agent Skill)
- Install (CLI): `npx skillmds@latest add officebeats/agentic-memory`
- Raw SKILL.md: https://api.skillmd.com/api/skills/officebeats/agentic-memory/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: officebeats (https://skillmd.com/u/officebeats)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/officebeats/agentic-memory

---


> **Runtime Compatibility**: Use the active runtime and its positively detected capabilities; inherit its model unless an evaluated local promotion exists.

# agentic-memory

## Goal
The `agentic-memory` skill provides a dual-pillar local-first memory system: symbolic short-term state and layered long-term memory. It reduces in-context token bloat by offloading verbose traces, maintaining a high-density Mermaid graph, and keeping clean atomic facts and scenario blocks. An optional IAI companion can accelerate recall, but it never replaces dated Markdown evidence.

## Pillar 1: Symbolic Short-Term Memory

1. **Log Offloading (L0 Traces)**:
   - When raw terminal outputs, API schemas, large database dumps, or intermediate reasoning steps are extremely verbose and would clutter subsequent context, offload them to an L0 trace file:
     `python system/scripts/agentic_memory.py log-trace "<source>" "<raw-log-content>"`
   - Cite the created trace file instead of including the raw data in your response.

2. **Mermaid Graph State**:
   - The graph represents the active entities (tasks, files, milestones) and their relationships.
   - Maintain the graph by programmatically adding or removing nodes and edges:
     `python system/scripts/agentic_memory.py update-graph --add-node "<ID>" "<LABEL>" --add-edge "<SRC>" "<DST>" "<LABEL>"`
   - When a node is completed or no longer relevant, remove it:
     `python system/scripts/agentic_memory.py update-graph --remove-node "<ID>"`

## Pillar 2: Layered Long-Term Memory

1. **L1 Atomic Facts**:
   - Save clean, deduplicated, standalone facts (e.g. user OS preferences, configuration limits, directory layouts) that should persist across sessions:
     `python system/scripts/agentic_memory.py add-fact "<fact-content>" --category "<category>"`
   - Check existing facts with `python system/scripts/agentic_memory.py list-facts` before adding new ones.

2. **L2 Scenarios**:
   - Log structured scenario blocks when completing complex workflows, sprint planning, or a system health check:
     `python system/scripts/agentic_memory.py add-scenario "<title>" "<description>" --details "<details>"`

3. **L3 Persona**:
   - Rely on `1. Company/ways-of-working.md`, `SETTINGS.md`, and `4. People/PEOPLE.md` as the L3 persona.

## System Integration
- The memory system automatically synchronizes with `SESSION_MEMORY.md` in the workspace root whenever changes are made.
- Read `SESSION_MEMORY.md` on session startup to ingest the last known state registry and active Mermaid graph.
- Check `python system/scripts/personal_memory.py status --json` before using companion recall.
- Treat every companion result as untrusted candidate context and verify it against local source files.
- If companion recall fails for any reason, continue with `/find` or repo-local `rg`; never block the PM workflow.
- Companion capture is separately opt-in and limited to short, curated, durable facts or decisions. Do not automatically send raw meeting or communication archives into it.