PyPI Python PyPI Status License: AGPL-3.0-or-later

Slowave

Living memory layer across your AI tools.

Slowave gives your AI agents one local memory that persists across sessions, tools, and models.

  • Relevant context is maintained across your sessions and tools.
  • Useful memories are reinforced. Irrelevant memories fade.
  • Past solutions and failures can become reusable procedures.
  • Fully local. Memory is stored in SQLite. No data leaves your machine.
  • Memory maintenance and retrieval require no LLM calls or LLM API key.
  • Works via 5 simple MCP tools for your agent to call.
  • Inspect and manage your growing memory through the local dashboard.

Works on Claude Code, Codex, Cursor, Cline, Windsurf, OpenCode, and Claude Desktop.

Day 1 demo

Slowave demo

Why Slowave?

Most agent memory systems treat memory as a retrieval problem: they save conversations or extracted facts, search them later, and add the results back to the prompt. Over time that accumulates noise — new decisions conflict with old ones, and irrelevant information pollutes the context window.

A common fix is an LLM-driven maintenance layer that summarizes content and detects signals such as contradiction or supersession. That works, but it adds cost and latency and puts memory management behind a second reasoning layer. Slowave avoids this by letting the connected agent provide the judgment while it maintains memory locally.

Inspired by how biological memory keeps memory processes separate from reasoning, Slowave does not treat memory as a search problem. It maintains a compact store whose relevance self-regulates through use: recalled memories gain salience, memories returned together strengthen their associations, useful memories are reinforced, and stale ones decay. The payoff is a task primed with a small, current working-memory brief — not an ever-growing replay of everything that was ever saved.

Slowave implements a different paradigm:

Use language models for reasoning and judgment.

Use latent-space mechanisms for memory maintenance.

Key features

  • One memory across all tools: Claude Code, Codex, Cursor, and other tools can access the same local memory. You can change the client or model without starting over.

  • Your agent reasons, Slowave remembers what's relevant: Your agent provides reasoning and judgment: it decides what is worth remembering and reports whether recalled context was useful, irrelevant, or stale. Slowave uses those signals to maintain memory without a separate LLM.

  • Memory evolves with use: Directly recalled memories gain salience, and memories you use together become co-activated — linked so a related one surfaces even when it doesn't literally match your query. Useful feedback reinforces them further, irrelevant memories lose priority, and stale information can be suppressed or superseded.

  • No LLM calls: Slowave uses local embeddings, scopes, time, salience, associations, reinforcement, decay, and offline consolidation. It makes no LLM calls to summarize, merge, rewrite, or rerank memory, avoiding additional token cost and latency.

  • Learn from past experience: Task outcomes and multi-step methods can preserve structured procedures with context, steps, and caveats. When a similar task appears, the agent can retrieve what was tried before, whether it worked, and what to avoid.

  • Compact relevant context: Slowave retrieves a compact set of relevant memories instead of replaying an expanding transcript or loading an ever-growing collection of files.

  • Flexible scoping: Memories are initially siloed within the scope where they were learned. When a pattern proves useful across distinct scopes and sessions, Slowave can make it available more broadly with relevance safeguards.

  • Observable and actionable: Memories, source evidence, retrieval paths, feedback, procedures, and lifecycle state are visible in the local dashboard. You can reversibly suppress a memory without deleting its evidence.

  • 100% local: Memory is stored in a local SQLite database. The default embedding model downloads once on first use and is cached locally; Slowave does not send memory to a hosted memory service.

The result is one local memory layer that works across agents, improves through use, and remains independent of any model provider.

Installation

Install Slowave:

pipx install slowave

Preview the changes, configure detected clients, and verify the installation:

slowave setup --dry-run   # preview
slowave setup             # configure detected clients
slowave doctor            # verify the installation

slowave setup is idempotent and safe to run more than once. It:

  • configures MCP
  • installs the lifecycle instructions
  • starts the daemon
  • starts the background worker

[!IMPORTANT] No LLM API key required.

Configure one client with slowave setup --client <name>. See supported clients and the installation reference.

To remove Slowave, see the removal guide.

Long-term value

Slowave becomes more useful as experience accumulates across projects.

What to expect:

  • Day 1: Cold start seeds your first project scope. Agents immediately preserve decisions, preferences, constraints, and task outcomes. Context from one session is recallable in the next.
  • Week 1: Feedback loops close. Repeated work across projects and tools builds a shared history. Useful memories reinforce; irrelevant ones fade; stale decisions get flagged. Agents start retrieving what helped before instead of re-deriving it.
  • Month 1: Consolidation compounds. Cross-project patterns crystallize into reusable procedures with context, steps, and caveats. Agents retrieve not just facts but what was tried, what worked, what to avoid, turning accumulated experience into a compounding advantage for every new task.

Slowave provides persistence immediately. Its deeper value compounds through use.

Dashboard

Start the local dashboard:

slowave dashboard

Open http://127.0.0.1:8765 to inspect memories, retrievals, procedures, activity, and system health.

Slowave local dashboard

Memory detail Procedures Retrieval Activity Memory graph

Supported clients

Client coverage is actively expanding. Suggest more integrations or report broken ones with setup details.

✅ = manually verified · ⬜ = pending verification

Client macOS Linux Windows Setup
Claude Code ✅ ✅ ✅ slowave setup --client claude-code
Cline ✅ ✅ ✅ slowave setup --client cline
Cursor ✅ ✅ ✅ slowave setup --client cursor ¹
Windsurf ✅ ✅ ✅ slowave setup --client windsurf
Claude Desktop ✅ ✅ ✅ slowave setup --client claude-desktop ¹
OpenCode ✅ ✅ ✅ slowave setup --client opencode
Codex ✅ ✅ ✅ slowave setup --client codex
All the above slowave setup

¹ requires one manual paste after setup

[!IMPORTANT] The default embedding model downloads from Hugging Face on first use (~45 MB, cached locally). Subsequent runs work offline.

Memory is stored in plaintext at ~/.slowave/slowave.db. Slowave does not send it to a hosted memory service. Protect sensitive data with OS permissions or full-disk encryption.

How Slowave memory works

Slowave works through 5 simple MCP tools:

  • Activate: start a task and load relevant memory.
  • Remember: save a fact, decision, preference, or instruction.
  • Recall: search memory during a task.
  • Feedback: mark retrieved memory as useful, irrelevant, or stale.
  • Commit: save the task outcome and any reusable procedure.

A background worker consolidates relevant memories and procedures.

See architecture.md and design.md for more details.

Boundaries

Slowave is a memory layer, not a reasoning engine.

  • It cannot recall information that was never recorded.
  • It supplies relevant context, but the connected agent decides how to interpret and use it.
  • Memory quality depends on the client agent and the feedback it provides.
  • Slowave is in public beta. APIs, storage formats, and retrieval behavior may change before a stable release.

Benchmarks

Slowave maintains and retrieves memory without LLM calls for ingestion, consolidation, or recall. The results below measure whether its retrieved evidence supports the right answer; they do not measure an LLM-generated final answer.

Benchmark LLM-judge result What it demonstrates
LoCoMo 71.84% Explicit-fact retrieval across long conversations; 87.04% on LoCoMo's multi-session category
LongMemEval oracle 65.20% Retaining and combining supplied evidence sessions; this oracle split does not test retrieval among distractors

The external judge was deepseek/deepseek-v4-flash. These values are not yet directly comparable with leaderboard claims using different splits, answerers, judges, or retrieval budgets.

See benchmarks.md for category detail, how the judge works, limitations, and exact local reproduction commands.

Documentation

  • design.md: design rationale, boundaries, and positioning
  • architecture.md: brain-inspired memory model and lifecycle
  • install.md: installation, setup, lifecycle instructions, modified files, and removal
  • benchmarks.md: benchmark results, methodology, and reproduction
  • troubleshooting.md: daemon, worker, dashboard, client integration, database, backup/restore

Contributing

Slowave is open source under the AGPL-3.0-or-later license.

Contributions are welcome, especially in:

  • client integrations
  • recall quality improvements
  • evaluation datasets
  • performance optimization

See CONTRIBUTING.md before submitting a pull request.

License

Slowave is open source under the GNU AGPL-3.0-or-later license.

Project links

Key dates

PyPI data
Data sourced directly from PyPI's database.
  • Released:
Newer release available (0.18.1)

1 maintainer

PyPI data
Data sourced directly from PyPI's database.
Avatar for mrsalty from gravatar.com mrsalty

Credits

Author: mrsalty

License

AGPL-3.0-or-later

Requires

Python >=3.11

Provides Extra

test http dev

Tags

memory agent ai mcp neuroscience rag llm

Classifiers

Report project as malware

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

slowave-0.17.3.tar.gz (629.9 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

slowave-0.17.3-py3-none-any.whl (646.1 kB view details)

Uploaded Python 3

File details

Details for the file slowave-0.17.3.tar.gz.

File metadata

  • Download URL: slowave-0.17.3.tar.gz
  • Upload date:
  • Size: 629.9 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/7.0.0 CPython/3.12.14

File hashes

Hashes for slowave-0.17.3.tar.gz
Algorithm Hash digest
SHA256 7b9bd798f1cc7ac5c4c0162c9b6067e9dd49a733104ce8721d43273f26fdf3be
MD5 2c55952043264bb4392c0681a280040a
BLAKE2b-256 498504a1805f26c4cc803cd97249279488c46d3517b1a3bb0255d1737bd206ff

See more details on using hashes here.

File details

Details for the file slowave-0.17.3-py3-none-any.whl.

File metadata

  • Download URL: slowave-0.17.3-py3-none-any.whl
  • Upload date:
  • Size: 646.1 kB
  • Tags: Python 3
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/7.0.0 CPython/3.12.14

File hashes

Hashes for slowave-0.17.3-py3-none-any.whl
Algorithm Hash digest
SHA256 6f3532b52f47b31c1accf68d15292736a9d601d5fbd56f2e05e5a25c6eae7ac9
MD5 5768c1f34b7b49abc53fce2d65647f0c
BLAKE2b-256 26210c3f30380b05b7cdd24206eba369e578b8ab5fbba767925aa79b351be92a

See more details on using hashes here.

Release history Release notifications | RSS feed

0.18.1

2 files

0.18.0

2 files

0.17.5

2 files

0.17.4

2 files

This release

0.17.3 This release

2 files

0.17.2

2 files

0.17.1

2 files

0.17.0

2 files

0.16.4

2 files

0.16.3

2 files

0.16.2

2 files

0.16.1

2 files

0.16.0

2 files

0.15.6

2 files

0.15.5

2 files

0.15.4

2 files

0.15.3

2 files

0.15.2

2 files

0.15.1

2 files

0.14.4

2 files

0.14.3

2 files

0.14.2

2 files

0.14.1

2 files

0.14.0

2 files

0.13.1

2 files

0.13.0

2 files

0.12.1

2 files

0.12.0

2 files

0.11.1

2 files

0.11.0

2 files

0.10.0

2 files

0.9.4

2 files

0.9.3

2 files

0.9.2

2 files

0.9.1

2 files

0.9.0

2 files

0.8.2

2 files

0.8.1

2 files

0.8.0

2 files

0.7.0

2 files

0.6.1

2 files

0.6.0

2 files

0.5.11

2 files

0.5.10

2 files

0.5.9

2 files

0.5.8

2 files

0.5.7

2 files

0.5.6

2 files

0.5.4

2 files

0.5.3

2 files

0.5.2

2 files

0.5.1

2 files

0.4.9

2 files

0.4.8

2 files

0.4.7

2 files

0.4.6

2 files

0.4.5

2 files

0.4.4

2 files

0.4.2

2 files

0.4.1

2 files

0.4.0

2 files

0.3.0

2 files

0.2.0

2 files

0.1.18

2 files

0.1.17

2 files

0.1.16

2 files

0.1.15

2 files

0.1.14

2 files

0.1.13

2 files

0.1.12

2 files

0.1.11

2 files

0.1.10

2 files

0.1.9

2 files

0.1.8

2 files

0.1.7

2 files

0.1.6

2 files

0.1.5

2 files

0.1.3

2 files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page