observational-memory 0.2.0

Cross-agent observational memory for Claude Code and Codex CLI

Observational Memory

Cross-agent shared memory for Claude Code and Codex CLI, with built-in search and no database setup.

Observational Memory runs two background jobs (Observer and Reflector) that condense transcript history into shared markdown memory. On session start, it retrieves relevant context with a pluggable search layer: built-in BM25 by default, plus optional QMD and qmd-hybrid backends for stronger semantic recall.

Adapted from Mastra's Observational Memory pattern. See the OpenClaw version for the original.

---

Quick start

Prerequisites

• Python 3.11+
• uv (recommended) or pip
• One LLM access path:
  • Direct API key (ANTHROPIC_API_KEY or OPENAI_API_KEY)
  • Google Vertex AI auth (ADC) for Anthropic on Vertex
  • AWS credentials/profile/role for Anthropic on Bedrock
• Claude Code and/or Codex CLI installed

Install

# Option A: Install from PyPI
uv tool install observational-memory

# Option A2: Install with enterprise provider dependencies
uv tool install "observational-memory[enterprise]"

# Option B (macOS): Install from Homebrew tap
brew tap intertwine/tap
brew install intertwine/tap/observational-memory

# Set up hooks, LLM provider config, and cron
om install

Verify

om doctor

That's it. Your agents now share persistent, compressed memory.

---

Why

If you use Claude Code in one terminal and Codex CLI in another, context gets lost fast. Each session starts cold.

Observational Memory fixes this. A single set of compressed memory files lives at ~/.local/share/observational-memory/ and is shared across all your agents:

Five tiers of memory

Tier	Updated	Retention	Size	Contents
Raw transcripts	Real-time	Session only	~50K tokens/day	Full conversation
Auto-memory	Hourly scan (no LLM)	Mirrors source	Per-project	Claude Code per-project discrete facts
Observations	Per session + periodic checkpoints (~15 min default)	7 days	~2K tokens/day	Timestamped, prioritized notes
Reflections	Daily	Indefinite	200–600 lines total	Durable long-term memory
Startup profile/act	Derived on install + observe/reflect	Derived	small startup slice	Compact default context for session start

---

How it works

Claude Code integration

SessionStart hook: On session start, om context injects compact derived startup files (profile.md + active.md) via additionalContext. If those files are missing, it regenerates them from reflections/observations. If om is unavailable, the shell fallback still supports the older full-file dump behavior.

SessionEnd hook: When a session ends, the observer runs on that transcript and compresses it into observations.

UserPromptSubmit / PreCompact hooks: Long sessions also trigger periodic checkpoints. They are throttled by OM_SESSION_OBSERVER_INTERVAL_SECONDS (default 900), so capture stays incremental without running on every prompt.

To disable in-session checkpoints while keeping normal end-of-session capture, set: OM_DISABLE_SESSION_OBSERVER_CHECKPOINTS=1 in ~/.config/observational-memory/env.

All hooks are installed automatically to ~/.claude/settings.json.

Claude Code auto-memory integration

Auto-memory as input source: Claude Code stores per-project discrete facts (preferences, feedback, decisions) in ~/.claude/projects/*/memory/*.md. The om observe --source claude-memory command scans these files, detects changes via content hashing, and indexes them into the search layer. Unlike transcript-based sources, auto-memory files are already distilled — they bypass the observer LLM entirely.

Cross-project enrichment: Auto-memory facts from all projects are supplied to the reflector as supplementary context, so knowledge from one project can surface when working in another.

Hourly cron: An hourly cron job runs the auto-memory scan (no LLM calls — just hash comparison and reindex). Deletions are detected and the reflector is instructed to clean up stale facts.

Codex CLI integration

AGENTS.md: The installer adds instructions to ~/.codex/AGENTS.md so Codex reads compact startup memory files at session start, with deeper memory available on demand via om search or the raw memory files.

Cron observer: A cron job runs every 15 minutes, scans ~/.codex/sessions/ for new transcript data (*.json and *.jsonl), and compresses it into observations. At the end of each observer run, om also checks whether reflections have fallen behind and runs a reflector catch-up when needed.

Reflector (both)

A daily cron job (04:00 UTC) runs the reflector, which:

1. Reads the Last reflected timestamp from the existing reflections
2. Filters observations to only those from that date onward (incremental; skips already-processed days)
3. If the filtered observations fit in one LLM call (<30K tokens), processes them in a single pass
4. If they're too large (e.g., after a backfill), automatically chunks by date section and folds each chunk into the reflections incrementally
5. Merges, promotes (🟡→🔴), demotes, and archives entries
6. Stamps Last updated and Last reflected timestamps programmatically
7. Writes the updated reflections.md
8. Trims observations older than 7 days

If that daily run is missed, for example because a laptop is asleep, the next successful om observe run will automatically catch reflections up to the newest observation date.

Priority system

Level	Meaning	Examples	Retention
🔴	Important / persistent	User facts, decisions, project architecture	Months+
🟡	Contextual	Current tasks, in-progress work	Days–weeks
🟢	Minor / transient	Greetings, routine checks	Hours

LLM providers and auth

The observer and reflector call an LLM API for compression. Provider and auth settings are stored in:

~/.config/observational-memory/env

om install creates this file with 0600 permissions (owner-read/write only). It supports both interactive setup and non-interactive flags.

Supported provider profiles:

Profile	OM_LLM_PROVIDER	Auth mode	Required settings
Direct Anthropic	anthropic	API key	ANTHROPIC_API_KEY
Direct OpenAI	openai	API key	OPENAI_API_KEY
Anthropic on Vertex	anthropic-vertex	Google ADC	OM_VERTEX_PROJECT_ID, OM_VERTEX_REGION
Anthropic on Bedrock	anthropic-bedrock	AWS credential chain	OM_BEDROCK_REGION (or AWS_REGION)
Legacy auto-detect	auto	API key	prefers ANTHROPIC_API_KEY, then OPENAI_API_KEY

The CLI, hooks, and cron jobs source this file automatically. You do not need to export keys in your shell profile.

Model selection precedence:

1. OM_LLM_OBSERVER_MODEL / OM_LLM_REFLECTOR_MODEL
2. OM_LLM_MODEL
3. Provider default (claude-sonnet-4-5-20250929 for Anthropic profiles, gpt-4o-mini for OpenAI)

Example direct key setup:

OM_LLM_PROVIDER=anthropic
ANTHROPIC_API_KEY=sk-ant-...

Example Vertex setup:

OM_LLM_PROVIDER=anthropic-vertex
OM_VERTEX_PROJECT_ID=my-gcp-project
OM_VERTEX_REGION=us-east5
OM_LLM_MODEL=claude-sonnet-4-5-20250929

Example Bedrock setup:

OM_LLM_PROVIDER=anthropic-bedrock
OM_BEDROCK_REGION=us-east-1
OM_LLM_MODEL=anthropic.claude-sonnet-4-5-20250929-v1:0

---

CLI reference

# Run observer on all recent transcripts
om observe

# Run observer on a specific transcript
om observe --transcript ~/.claude/projects/.../abc123.jsonl

# Run observer for one source only
om observe --source claude
om observe --source codex
om observe --source claude-memory

# Run reflector
om reflect

# Search memories
om search "PostgreSQL setup"
om search "current projects" --limit 5
om search "backfill" --json
om search "preferences" --reindex   # rebuild index before searching

# Backfill all historical transcripts
om backfill --source claude
om backfill --dry-run               # preview what would be processed

# Dry run (print output without writing)
om observe --dry-run
om reflect --dry-run

# Install/uninstall
om install [--claude|--codex|--both] [--no-cron]
om install --provider anthropic-vertex --vertex-project-id my-proj --vertex-region us-east5 --llm-model claude-sonnet-4-5-20250929 --non-interactive
om install --provider anthropic-bedrock --bedrock-region us-east-1 --llm-model anthropic.claude-sonnet-4-5-20250929-v1:0 --non-interactive
om uninstall [--claude|--codex|--both] [--purge]

# Check status
om status

# Run diagnostics
om doctor
om doctor --json              # machine-readable output
om doctor --validate-key      # test configured provider access with a live call

---

Configuration

LLM provider settings

~/.config/observational-memory/env

Created by om install with 0600 permissions. Typical values:

OM_LLM_PROVIDER=anthropic
OM_LLM_MODEL=claude-sonnet-4-5-20250929
ANTHROPIC_API_KEY=sk-ant-...

This file is sourced by the om CLI, the Claude Code hooks, and the cron jobs. Environment variables already present in your shell take precedence.

Memory location

Default: ~/.local/share/observational-memory/

Key files:

• profile.md — compact stable startup profile
• active.md — compact active startup context
• reflections.md — full long-term memory
• observations.md — recent detailed notes

Override with XDG_DATA_HOME:

export XDG_DATA_HOME=~/my-data
# Memory will be at ~/my-data/observational-memory/

Cron schedules

The installer sets up:

• Observer (Codex): */15 * * * * by default (controlled by OM_CODEX_OBSERVER_INTERVAL_MINUTES, e.g. */10 * * * * for 10 min)
• Auto-memory scan: 0 * * * * (hourly, no LLM calls — just hash comparison and reindex)
• Reflector: 0 4 * * * (daily at 04:00 UTC)

Set OM_CODEX_OBSERVER_INTERVAL_MINUTES in ~/.config/observational-memory/env to tune Codex polling (1 = every minute).

Edit with crontab -e to adjust.

Search backend

Memory search uses a pluggable backend architecture. Three backends are available:

Backend	Default	Requires	Method
bm25	Yes	Nothing (bundled)	Token-based keyword matching via rank-bm25
qmd	No	QMD CLI + bun	BM25 keyword search via QMD's FTS5 engine
qmd-hybrid	No	QMD CLI + bun	Hybrid BM25 + vector embeddings + LLM reranking (~2GB models, auto-downloaded)
none	No	Nothing	Disables search entirely

The default bm25 backend works out of the box. The index is rebuilt automatically after each observe/reflect run and stored at ~/.local/share/observational-memory/.search-index/bm25.pkl.

To switch backends, set OM_SEARCH_BACKEND in your env file:

# ~/.config/observational-memory/env
OM_SEARCH_BACKEND=qmd-hybrid
OM_CODEX_OBSERVER_INTERVAL_MINUTES=10

Or export it in your shell:

export OM_SEARCH_BACKEND=qmd-hybrid
export OM_CODEX_OBSERVER_INTERVAL_MINUTES=10

Using QMD (optional)

QMD provides hybrid search (BM25 + vector embeddings + LLM reranking) for better recall on semantic queries. Models run locally through node-llama-cpp, so no extra API key is required. To set it up:

# 1. Install bun (QMD runtime)
curl -fsSL https://bun.sh/install | bash

# 2. Install QMD (from GitHub — the npm package is a placeholder)
bun install -g github:tobi/qmd

# 3. Switch the backend in config.py
#    search_backend: str = "qmd-hybrid"

# 4. Rebuild the index
om search --reindex "test query"

When using QMD, memory documents are written as .md files under ~/.local/share/observational-memory/.qmd-docs/. They are registered as a QMD collection named observational-memory. om search and om context use whichever backend is configured.

Tuning

Edit the prompts in prompts/ to adjust:

• What gets captured: priority definitions in observer.md
• How aggressively things are merged: rules in reflector.md
• Target size: the reflector aims for 200 to 600 lines

---

Example output

Observations (observations.md)

# Observations

## 2026-02-10

### Current Context

- **Active task:** Setting up FastAPI project for task manager app
- **Mood/tone:** Focused, decisive
- **Key entities:** Atlas, FastAPI, PostgreSQL, Tortoise ORM
- **Suggested next:** Help with database models

### Observations

- 🔴 14:00 User is building a task management REST API with FastAPI
- 🔴 14:05 User prefers PostgreSQL over SQLite for production (concurrency)
- 🟡 14:10 Changed mind from SQLAlchemy to Tortoise ORM (finds SQLAlchemy too verbose)
- 🔴 14:15 User's name is Alex, backend engineer, prefers concise code examples

Reflections (reflections.md)

# Reflections — Long-Term Memory

_Last updated: 2026-02-10 04:00 UTC_
_Last reflected: 2026-02-10_

## Core Identity

- **Name:** Alex
- **Role:** Backend engineer
- **Communication style:** Direct, prefers code over explanation
- **Preferences:** FastAPI, PostgreSQL, Tortoise ORM

## Active Projects

### Task Manager (Atlas)

- **Status:** Active
- **Stack:** Python, FastAPI, PostgreSQL, Tortoise ORM
- **Key decisions:** Postgres for concurrency; Tortoise ORM over SQLAlchemy

## Preferences & Opinions

- 🔴 PostgreSQL over SQLite for production
- 🔴 Concise code examples over long explanations
- 🟡 Tortoise ORM over SQLAlchemy (less verbose)

---

Contributing and maintainers

Contributor and maintainer instructions have moved to docs/MAINTAINERS.md.

How it compares to the OpenClaw version

Feature	OpenClaw Version	This Version
Agents supported	OpenClaw only	Claude Code + Codex CLI
Scope	Per-workspace	User-level (shared across all projects)
Observer trigger	OpenClaw cron job	Claude: SessionEnd hook; Codex: system cron
Context injection	AGENTS.md instructions	Claude: SessionStart hook; Codex: AGENTS.md
Memory location	workspace/memory/	~/.local/share/observational-memory/
Compression engine	OpenClaw agent sessions	Direct LLM API calls (Anthropic/OpenAI)
Cross-agent memory	No	Yes

---

FAQ

Q: Does this replace RAG / vector search? A: For personal context, mostly yes. Observational memory tracks facts about you (preferences, projects, working style). RAG is still better for large document collections. Use BM25 for lightweight local retrieval, or qmd-hybrid with QMD if you want hybrid semantic search.

Q: How much does it cost? A: The observer processes only new messages per session (~200–1K input tokens typical). The reflector runs once daily. Expect ~$0.05–0.20/day with Sonnet-class models.

Q: What if I only use Claude Code? A: Run om install --claude. The Codex integration is entirely optional.

Q: Can I manually edit the memory files? A: Yes. Both observations.md and reflections.md are plain markdown. The observer appends; the reflector overwrites. Manual edits to reflections will be preserved.

Q: What happens if the reflector runs on a huge backlog? A: The reflector runs incrementally. It reads Last reflected from reflections.md and only processes newer observations. If that timestamp is missing (first run or after backfill), it chunks observations by date and folds them in batches so the model is not overloaded. Output budget is 8192 tokens, which is enough for the 200 to 600 line target.

Q: What about privacy? A: Everything runs locally. Transcripts are processed by the LLM API you configure (Anthropic or OpenAI), subject to their data policies. No data is sent anywhere else.

---

Credits

• Inspired by Mastra's Observational Memory
• Original OpenClaw version
• License: MIT