slowave 0.4.4

Brain-inspired long-term memory for AI agents — zero LLM during ingest or retrieval

Slowave

A second brain for your AI, shared across every tool.

Slowave gives your AI private, local memory that persists across sessions, evolves over time, and costs nothing to run — no API key, no LLM calls, no data leaving your machine.

Why Slowave?

👊 One memory, every AI tool.
Claude Code, Cline, Claude Desktop, Cursor, Windsurf, and any MCP-compatible client share the same local memory store. Fix a bug in Claude Code tonight — Cline knows the lesson tomorrow. Decide on an architecture in Claude Desktop — it surfaces in your next coding session. Context follows you across tools instead of dying when you close a chat.

🧠 Adaptive memory, not static notes.
Most AI memory is a pile of Markdown. Slowave behaves more like a brain: frequently recalled memories strengthen, stale ones fade, contradicted facts get superseded automatically. You never manually clean up a MEMORY.md file again.

⚙️ Procedural memory: workflows that stick.
Slowave stores reusable procedures — "how we do deploys in this repo", "steps to implement a new feature across projects" or simply "how this spaghetti recipe should be cooked". Recall them by goal and situation, not by keyword search. Your agents learn habits, not just facts.

🔒 Fully local, zero LLM calls.
Ingestion, consolidation, and recall run on your machine using embeddings, FAISS, and SQLite — no LLM in the memory loop, no API key, no data sent to a cloud memory backend. Memory operations cost $0 per query and work offline.

💰 86% fewer tokens than replaying history.
Slowave injects a compact working-memory brief instead of accumulating the full conversation. Over 20 sessions, raw history grew from 96 → 1,875 tokens while Slowave stayed flat at ~136 tokens — with 95% recall quality (the right memory surfaced in 19/20 sessions). Crossover happens at session 2. Measured with the real semantic encoder, not claimed. See the test →

Install

pipx install slowave
slowave setup           # automated configuration

slowave setup detects your platform, wires every client it finds, injects lifecycle hooks, and starts the background worker. Idempotent and safe to re-run. See what gets modified →

Uninstall:

slowave cleanup         # remove all configuration
pipx uninstall slowave  # remove package

[!IMPORTANT] Claude Desktop: after setup, paste the lifecycle block into Settings → General → Instructions for Claude. Cursor: after setup, paste the lifecycle block into Settings → Rules for AI. slowave setup prints the exact text and location for both. All other clients (Cline, Claude Code, Windsurf) are fully automated.

slowave doctor   # verify installation
slowave stats    # memory snapshot

Memory is stored at ~/.slowave/slowave.db. No Ollama, no vector database, no cloud service required.

Full install guide →

What Slowave remembers

Anything that should survive across sessions: preferences, decisions, constraints, lessons learned, open questions, and reusable workflows — for work, research, or personal use. Each memory carries a timestamp, decays if never recalled, and strengthens when it proves useful. Contradictions are detected geometrically and old facts are superseded automatically — no LLM required.

Memory is scoped flexibly: project:my-app, domain:cooking, relationship:alex — or unscoped for universal context.

Benchmarks

Alpha-stage numbers. Internal runs, not independently verified. See docs/benchmarks.md for per-category results, ablation details, and known gaps.

Numbers from the clean ablation sweep (17 variants, strict sample-size validation, zero LLM calls throughout). Full system = salience reranking + episodic consolidation enabled.

Benchmark	n	Cosine baseline	Full system	Δ	LLM calls
LongMemEval	500	87.6%	87.8%	+0.2 pp (saturation)	0
LoCoMo	1986	72.1%	83.5%	+11.4 pp	0
DMR (MSC Self-Instruct)	500	93.6%	88.0%	−5.6 pp ⚠	0
StaleMemory (concrete attrs)	900	—	86–89% detection	—	0

⚠ DMR note: The full system's −5.6 pp on DMR is a protocol artefact — DMR uses keyword-overlap scoring, which penalises salience reranking and schema consolidation (abstractions that improve recall on conversational queries but reduce raw keyword matches). The cosine-only baseline (93.6%) is the fair DMR headline.

Documentation

docs/design	the brain-inspired rationale behind Slowave
docs/architecture.md	How memory consolidation works
docs/install.md	Install, setup, per-client wiring, troubleshooting
docs/slowave_setup.md	slowave setup command help
docs/manual_setup.md	Step-by-step manual configuration guide
docs/benchmarks.md	Per-category results, known gaps, reproducibility
docs/token_efficiency.md	Token efficiency vs. history replay and static knowledge files
docs/limitations.md	Honest limits: scale, language, unsolved categories
docs/cli.md	CLI reference
docs/dashboard.md	Local web UI (slowave dashboard)

Dashboard

Keep your second-brain always under control through the local dahsboard.

You use it, your second-brain will start connecting the dots

Contributing

Slowave is open source under AGPL-3.0-or-later. Bug reports, install feedback, and focused improvements are welcome — read CONTRIBUTING.md before opening a PR. Commercial licensing terms may be offered in the future.