Memory Pipeline

Memory Pipeline + Performance Routine

A complete memory and performance system for AI agents. Two subsystems, one package:

• Memory Pipeline (Python scripts) — Extracts facts, builds knowledge graphs, generates daily briefings
• Performance Routine (TypeScript hooks) — Pre-game briefing injection, tool discipline, output compression, after-action review

What This Does

Memory Pipeline (Between Sessions)

A three-stage system that helps AI agents maintain long-term memory:

1. Extract — Pulls structured facts (decisions, preferences, learnings, commitments) from daily notes and session transcripts using LLM extraction
2. Link — Builds a knowledge graph with embeddings and bidirectional links between related facts, identifies contradictions
3. Briefing — Generates a compact BRIEFING.md file loaded at session start with personality reminders, active projects, recent decisions, and key context

Performance Routine (Within Sessions)

Four lifecycle hooks that enforce consistency during agent runs:

1. Pre-Game Routine (before_agent_start) — Assembles a bounded briefing packet from memory files + checklist, injects into system prompt
2. Tool Discipline (before_tool_call) — Enforces deny lists, normalizes params, prevents unsafe tool calls
3. Output Compression (tool_result_persist) — Head+tail compression of large tool results to prevent context bloat
4. After-Action Review (agent_end) — Writes durable notes about what happened, tools used, and outcomes

Quick Start

Installation

The skill includes three Python scripts in scripts/:

• memory-extract.py — Fact extraction
• memory-link.py — Knowledge graph building
• memory-briefing.py — Daily briefing generation

All scripts auto-detect your workspace from:

1. CLAWDBOT_WORKSPACE environment variable
2. Current working directory (if contains SOUL.md or AGENTS.md)
3. ~/.clawdbot/workspace (default fallback)

Requirements

At least one LLM API key is required:

• OpenAI API key (for GPT-4o-mini + embeddings)
• Anthropic API key (for Claude Haiku)
• Gemini API key (for Gemini Flash)

Set via environment variable or config file:

# Environment variable
export OPENAI_API_KEY="sk-..."

# OR config file
echo "sk-..." > ~/.config/openai/api_key

The scripts will auto-detect and use whichever API key is available.

Basic Usage

Run the full pipeline:

python3 scripts/memory-extract.py
python3 scripts/memory-link.py
python3 scripts/memory-briefing.py

Or run individual steps as needed.

Pipeline Stages

Stage 1: Extract Facts

Script: memory-extract.py

Reads from (in priority order):

1. Daily memory files ({workspace}/memory/YYYY-MM-DD.md) — today or yesterday
2. Session transcripts (~/.clawdbot/agents/main/sessions/*.jsonl)

Extracts structured facts:

• Type: decision, preference, learning, commitment, fact
• Content: The actual information
• Subject: What it's about (auto-detected from context)
• Confidence: 0.0-1.0 reliability score

Output: {workspace}/memory/extracted.jsonl — One JSON fact per line, deduplicated

Stage 2: Build Knowledge Graph

Script: memory-link.py

Takes extracted facts and:

• Generates embeddings (if OpenAI key available, else uses keyword similarity)
• Creates bidirectional links between related facts
• Detects contradictions and marks superseded facts
• Auto-generates domain tags from content

Output:

• {workspace}/memory/knowledge-graph.json — Full graph with nodes and links
• {workspace}/memory/knowledge-summary.md — Human-readable summary

Stage 3: Generate Briefing

Script: memory-briefing.py

Creates a compact daily briefing loaded at session start.

Combines:

• Personality traits (from SOUL.md if exists)
• User context (from USER.md if exists)
• Active projects (top subjects from recent facts)
• Recent decisions and preferences
• Active todos (from any todos*.md files)

Output: {workspace}/BRIEFING.md — Under 2000 chars, LLM-generated or template-based

Wiring Into HEARTBEAT.md

To run automatically, add to your workspace's HEARTBEAT.md:

# Heartbeat Tasks

## Daily (once per day, morning)
- Run memory extraction: `cd {workspace} && python3 skills/memory-pipeline/scripts/memory-extract.py`
- Build knowledge graph: `cd {workspace} && python3 skills/memory-pipeline/scripts/memory-link.py`
- Generate briefing: `cd {workspace} && python3 skills/memory-pipeline/scripts/memory-briefing.py`

## Weekly (Sunday evening)
- Review `memory/knowledge-summary.md` for insights
- Clean up old daily notes (optional)

Loading BRIEFING.md

Important: BRIEFING.md needs to be loaded as workspace context at session start. This requires the OpenClaw context loading feature (currently in development).

Once available, configure your agent to load BRIEFING.md along with SOUL.md, USER.md, and AGENTS.md at the start of each session.

Output Files

All files are created in {workspace}/memory/:

• extracted.jsonl — All extracted facts (append-only)
• knowledge-graph.json — Full knowledge graph with embeddings and links
• knowledge-summary.md — Human-readable summary of the graph
• BRIEFING.md (in workspace root) — Daily context cheat sheet

Customization

Changing Models

Edit the model names in each script:

• memory-extract.py: Lines with "model": "gpt-4o-mini" (or claude/gemini equivalents)
• memory-link.py: Line with "model": "text-embedding-3-small"
• memory-briefing.py: Lines with "model": "gpt-4o-mini"

Adjusting Extraction

In memory-extract.py, modify the extraction prompt (lines ~75-85) to focus on different types of information or change the output format.

Link Threshold

In memory-link.py, change the similarity threshold for creating links (currently 0.3 at line ~195).

Troubleshooting

No facts extracted:

• Check that daily notes or transcripts exist
• Verify API key is set correctly
• Check script output for LLM errors

Low-quality links:

• Add OpenAI API key for embedding-based similarity (more accurate than keyword matching)
• Adjust similarity threshold in memory-link.py

Briefing too long:

• Reduce number of facts included in template (edit generate_fallback_briefing)
• LLM-generated briefings are automatically constrained to 2000 chars

Performance Routine (Hook System)

The performance routine is implemented as OpenClaw lifecycle hooks in src/. It applies a core principle from performance psychology: separate thinking from doing. Athletes don't redesign their technique mid-game — they prepare (purposeful thinking), then execute trained sequences (reactive execution). The only exception is genuine error handling.

For agents, this means: front-load all context, constraints, and memory retrieval into a briefing packet before inference starts. Keep execution clean. Write the after-action review after. Never inject corrections mid-run.

Architecture

User Message → Gateway → Agent Loop
  ├── before_agent_start → Briefing Packet (checklist + memory + constraints)
  ├── LLM Inference (clean context, no mid-run corrections)
  ├── before_tool_call → Policy enforcement (deny list)
  ├── Tool Execution → Result
  ├── tool_result_persist → Compression (head+tail, bounded)
  └── agent_end → After-Action Review → durable memory for next run

The Core Idea: No Mid-Swing Coaching

Constant correction during execution degrades output. Mid-run prompt patches create instruction collision — two competing directives the agent must reconcile instead of executing. The alternative:

1. Capture corrections — don't inject them into the current run
2. Condense into deltas — merge all corrections into a clean update
3. Inject next run — the next briefing packet includes the corrected instructions

The after-action review (agent_end) feeds back into the next briefing (before_agent_start). The loop is closed — just not during execution.

Configuration

Configure via openclaw.plugin.json or your agent config:

{
  "enabled": true,
  "briefing": {
    "maxChars": 6000,
    "checklist": [
      "Restate the task in one sentence.",
      "List constraints and success criteria.",
      "Retrieve only the minimum relevant memory.",
      "Prefer tools over guessing when facts matter."
    ],
    "memoryFiles": ["memory/IDENTITY.md", "memory/PROJECTS.md"]
  },
  "tools": {
    "deny": ["dangerous_tool"],
    "maxToolResultChars": 12000
  },
  "afterAction": {
    "writeMemoryFile": "memory/AFTER_ACTION.md",
    "maxBullets": 8
  }
}

Hook Details

before_agent_start — Briefing Packet

• Loads configured memory files from workspace
• Builds bounded packet: task hint + checklist + retrieved memory
• Injects into system prompt (respects maxChars limit)
• Missing memory files are silently skipped

before_tool_call — Tool Discipline

• Checks tool name against deny list
• Throws error if denied (prevents execution)
• Extensible for param normalization

tool_result_persist — Output Compression

• Keeps results under maxToolResultChars (default 12K)
• Uses head (60%) + tail (30%) strategy
• Preserves structure for JSON results

after_action_review — Durable Notes

• Appends session summary to configured memory file
• Extracts key bullets from final answer
• Logs tools used (with failure flags)
• Creates directories automatically

Source Files

• src/index.ts — Hook registration and wiring
• src/briefing.ts — Briefing packet builder
• src/compress.ts — Tool result compressor
• src/memory.ts — Memory file loader + after-action writer

See Also

• Setup Guide — Detailed installation and configuration
• Blog Post Draft — Full writeup of the performance routine concept