memory-context-pipeline

Built by DARKSOL 🌑

Local-model memory/context pipeline for OpenClaw that keeps premium reasoning on the main chat path while offloading summarization and memory extraction.

Why this exists

Long-running assistant sessions accumulate expensive context. This project adds an offload path so local workers can summarize, compress, and extract memory candidates while premium models stay focused on high-value user-facing reasoning.

What it does

• Adds a reusable OpenClaw skill package (skill/)
• Provides JSON output contract + validation scripts
• Supports auto-trigger checks around context-size thresholds
• Stores optional per-channel memory records
• Includes one-command orchestrator for trigger -> route -> conditional write
• Ships docs/examples for practical integration

Quickstart

# from repo root
python skill/scripts/validate_output.py --input examples/sample-output.json
python skill/scripts/validate_vision_tasks.py

# orchestrate one full cycle
python skill/scripts/orchestrate_memory_cycle.py --root . --guild-id <gid> --channel-id <cid> --task summarize --chars 25000

Real example(s)

# decide whether context should trigger memory flow
python skill/scripts/auto_trigger.py --guild-id <gid> --channel-id <cid> --chars 25000

# write channel memory record when trigger=true
python skill/scripts/channel_memory_store.py --guild-id <gid> --channel-id <cid> --text "Summarized key decisions"

Config/options

Script	Key args	Description
auto_trigger.py	--guild-id, --channel-id, --chars	Decides whether threshold conditions should trigger pipeline work
context_router.py	task/context args	Maps task class to local/premium routing decision
channel_memory_store.py	--guild-id, --channel-id, --text	Persists memory snippets into channel-scoped files
orchestrate_memory_cycle.py	--root, --guild-id, --channel-id, --task, --chars	End-to-end orchestration helper
validate_output.py	--input	Validates output against contract
enforcement_check.py	--config	Verifies policy wiring against OpenClaw config

Architecture / flow

• Trigger logic evaluates context size (auto_trigger.py)
• Router determines suitable execution path (context_router.py)
• Worker output is validated against JSON contract (skill/references/output_contract.json)
• Memory records are optionally written by channel/day (channel_memory_store.py)
• Orchestrator script can chain the full flow in one command

Performance notes

No benchmark numbers are claimed in this repo. Gains depend on your local model choice, hardware, and traffic shape.

Limitations + roadmap

Current limitations

• Focused on OpenClaw integration patterns, not a standalone hosted service
• Quality of summaries/extractions depends on local model behavior
• Threshold trigger is intentionally a soft guard, not a hard cutoff

Roadmap

• Broader policy templates for different deployment risk profiles
• More examples for mixed text/vision routing
• Additional end-to-end integration tests

Security notes

• Do not store secrets in memory files.
• Validate config before enabling automated writes in shared environments.

License + links

• License: MIT
• Architecture docs: docs/ARCHITECTURE.md
• Validation notes: docs/VALIDATION.md
• Release notes: docs/RELEASE_NOTES_*.md
• GitHub: https://github.com/darks0l/memory-context-pipeline