WhiteMagic Labs
Give your AI agent memory that survives restarts, model changes, and long projects.
WhiteMagic is our open-source, local-first memory layer for AI agents.
WhiteMagic Labs studies how AI agents remember, resume, coordinate and recover across long-running work.
curl -fsSL "https://www.whitemagic.dev/install.sh?ref=home-quick-install" | sh -s -- --version v9.3.4 --dir "$HOME/.local/bin" --ref home-quick-installAsk your AI assistant to set it up →·Hosted lanes & x402 Agent Payments →
continuity example
A decision recovered in a fresh process.
A fresh process recovered a saved choice, its reason, and the next step. This is one pinned, local demonstration; the conversation animation is illustrative.
Recovered after restart
“We chose cedar offcuts for the folding trellis because it fits the narrow bed and stores flat in winter.”
Next step: Measure the crosspieces
Recorded with v9.3.4 on Linux x86_64 musl: four separate CLI processes, a fresh throwaway store, and no connected model. The garden scenario is authored; this shows this release and path on one Linux host.
Inspect the four CLI steps and reproduction commands
Recorded terminal output · v9.3.4 · Linux x86_64 musl
Captured outputs & build identityWe ran four separate CLI processes against a fresh throwaway store. The garden scenario is authored; the recovered words and next step are actual output. No model was connected. This demonstrates this release and path on one Linux host.
Before reproducing: use the pinned release linked in the record, check its SHA-256 and wm --version, then use a new working directory. Run mkdir ./demo-store; if it already exists, stop and choose a new directory.
$ wm session continuity --store ./demo-store{
"briefing": {
"briefing_bytes": 258,
"briefing_hash": "7f8c38dbedca86f4e33bb68a0e2a9bbf1b03a4ec47e1dd0a59585b83ce054d57",
"estimated_tokens": 65,
"format_version": "wm-briefing/1",
"ordering": "checkpoint-hash, then turns oldest-to-newest",
"recent_verbatim": true,
"turn_hashes": [
{
"hash": "61d546b06fd131a6e80483e57c05ebb79b755ea1b4b455902310d665a9bb9104",
"memory_id": "c0df4288-109a-446c-8fdf-99e0fbee4f85"
}
]
},
"checkpoint": {
"next_queue": [
"Measure the crosspieces"
]
},
"checkpoint_id": "6f425dea-4419-4793-8ba3-27af9831af96",
"count": 1,
"max_content_bytes": 8192,
"max_response_bytes": 49152,
"previous_session": "26e39da7-c6b1-4a5a-9917-4ec7fcea6401",
"status": "success",
"truncated": false,
"turns": [
{
"content": "We chose cedar offcuts for the folding trellis because it fits the narrow bed and stores flat in winter.",
"content_bytes": 104,
"content_truncated": false,
"importance": 0.5,
"memory_id": "c0df4288-109a-446c-8fdf-99e0fbee4f85",
"role": "user",
"sequence": 1,
"session_id": "26e39da7-c6b1-4a5a-9917-4ec7fcea6401",
"turn_type": "decision"
}
],
"turns_omitted": 0
}The recovered result retains why cedar offcuts were chosen and the next step, “Measure the crosspieces.” Copy the commands to try it in your own disposable example directory.
View illustrative conversation animation
illustrative replay
▍
·
local core, independent lab
A small local product, built in public.
WhiteMagic Labs develops the open-source core and publishes evidence with clear limits. The local software is free to use; optional hosted services are separate and may carry charges.
WhiteMagic began in October 2025 with one person and a second-hand Dell. The shipped software now asks a narrower question: after months away, can an agent recover what it decided, and why?
channels and evidence
Separate channels, separate records.
Stable remains the default install choice. The experimental preview has its own version, binary, and store path. Each release record belongs to its own channel.
operating rules
Six rules the work answers to.
These decide the uncomfortable cases – what gets published, what ships, and what waits for evidence.
Label every claim
A measurement, an observation, a guess, and an ambition are four different objects. Each is tagged as what it is, and a claim that cannot point at its source is rewritten or dropped.
Keep the core on your hardware
The default install is a local binary and a local store. Networked conveniences may arrive later, but nothing in the core depends on them, and the local path is not a degraded demo.
Observation earns enforcement
Nothing that throttles, blocks, or isolates is trusted before it can show a measured false-positive rate. Watching first is a governance position, not just an engineering step.
Portability over lock-in
Memory sits in documented formats on your disk. Changing model, provider, editor, or machine is a configuration change — the history does not come along as a migration crisis.
Design for both sides of the table
People and agents will share this work. Usability, consent, attribution, and accountability are inputs to the design, not outside costs discovered after launch.
Publish the losses
Evaluations are built so the lab's own software can finish behind, and failures are documented next to wins. A result that cannot embarrass the lab is not treated as evidence.
questions, answered plainly
About the work and the local core.
A few common questions, with links to the detailed setup and product references.
In one paragraph, what is WhiteMagic?
An open-source continuity layer for AI agents: a local store for memories and sessions, a small MCP surface to reach them, governance records at the dispatch seam, and enough metadata to resume work after a restart, an upgrade, or a model switch. It runs on your hardware and needs no account.
Where does the data physically live?
In a local store under your home directory by default: LMDB for key-value persistence, Tantivy for full-text search, and memories organized into 16 galaxies – 15 cognition galaxies plus one that holds local telemetry evidence and is excluded from default recall. Stores are not encrypted by default; backups, exports, and replicas are separate copies, and deleting a store directory does not establish secure erasure.
Does anything leave my machine?
No memory, prompt, or session content is transmitted off-device, and no usage telemetry is sent off-device by default. Local diagnostic telemetry may be recorded on-device for self-observation; the only sharing path is the optional install-funnel report, explicit opt-in and off by default (`wm telemetry enable --share` prints the exact payload first; reset-id rotates its random identifier). Release/update checks contact GitHub only when run, and work offline. There is no account, and nothing needs configuring to keep working offline.
What is actually measured today?
CI test counts are unavailable in the current release manifest, and no count is asserted here; every public count on this site regenerates from the signed manifests. A published snapshot records microsecond-scale median dispatch on the Rust hot path – a routing figure, not end-to-end latency. Anything beyond that carries an Observed, Hypothesis, or Research goal label.
How do I connect an agent to it?
Stable Gen 2 v9.3.4: Linux arm64 (release manifest declares install gate); Linux arm64 musl (published asset); Linux x86-64 (release manifest declares install gate); Linux x86-64 musl (published asset); macOS arm64 (release manifest declares install gate); macOS x86-64 (release manifest declares install gate); Windows x86-64 (release manifest declares install gate). Experimental Gen 3 v10.3: Linux arm64 (release manifest declares install gate); Linux x86-64 (release manifest declares install gate); macOS arm64 (release manifest declares install gate). Gen 3 preview publishes no Windows or Intel macOS binary. The stable Install guide provides platform-specific commands, client configuration, and first-run checks: whitemagic.dev/install. Run wm grimoire to inspect readiness; wm connect previews client configuration, and --write applies it. The server speaks MCP locally over stdio.
join the conversation
Tell us what you find.
Share how WhiteMagic fits alongside the tools you already use. Useful combinations, surprises, and failed experiments are all welcome.
instant evaluation
Try hosted memory in seconds. Free evaluation keys.
No installation required. Generate an instant evaluation token in 1 click (50 recalls/day, no signup) to connect Claude Code, Cursor, Cline, or custom agents directly to WhiteMagic hosted recall.
Instant Free Evaluation Key (50 recalls/day · No Signup)
Generate a live Bearer token in 1 click. Connect Claude Code, Cursor, Cline, or agents instantly.
ready when you are
Start with the stable local core.
One command installs the stable channel. The Install guide has client setup, first-run checks, troubleshooting, and separate experimental preview instructions.