01
Sessions reset too easily
Without long-term memory, every new session starts blank. Preferences, project facts, and earlier decisions all disappear when you switch tools.
v3.9 release · A memory layer for AI agents
Memory Palace
Make AI remember with evidence, boundaries, and continuity.
Persistent. Searchable. Auditable.
Memory Palace gives every AI agent a memory that stays put — across sessions, across tools, across teammates. v3.9 makes that memory forget gracefully, retrieve smarter, and refuse silent overwrites. A dashboard, bilingual docs, and one-click Docker are already in the repo.
Remembers across sessions
Safe writes, easy rollback
Auto cleanup & layering
Smarter hybrid search
Run locally or scale up
>read_memory("system://boot")
>search_memory("coding style", include_session=true)
>create_memory("core://agent", ...)
>compact_context(force=false)Write Guard runs before create/update. Search, review, and maintenance stay on the same repo-local path.
v3.9 · 2026-05-20 validation snapshot
- Backend tests
- 1605 passed / 22 skipped
- Documented validation snapshot; see Docs Center.
- Frontend tests
- 232 passed
- Vitest rerun, plus typecheck and build passed.
Why this exists
AI agents still forget the moment you close the chat.
Memory Palace was built around three things that go wrong every day: forgetting, untrustworthy search, and writes you can't undo.
02
Search results you can't trust
Retrieval quality depends on mode, embeddings, and index health. Memory Palace shows you those knobs and gives you tools to fix them, instead of treating search as a black box.
03
Writes you can't take back
Memory shouldn't silently duplicate, overwrite, or drift. Every change passes through a guard and leaves a snapshot, so you can always see what happened and roll it back.
Memory is not a feature. It is the substrate that decides whether an agent stays consistent across sessions, tools, and teammates. Memory Palace design principle
Core flow
A repo-local workflow from recall to maintenance.
The public tool contract already documents a five-stage usage loop. This page compresses it into the practical path most users actually take.
Boot
Start with system://boot so the agent gets core memories and recent updates before it starts planning or writing.
Recall
Use search_memory and read_memory to find the right node first, instead of guessing paths or duplicating facts.
Write safely
Create and update operations pass through Write Guard, with duplicate detection, suggested targets, and snapshot-backed rollback boundaries.
Govern
Compact long sessions, rebuild indexes when retrieval drifts, and inspect runtime state through the dashboard and maintenance paths.
What's new in v3.9
What v3.9 actually shipped.
Six concrete improvements you can feel — already running in the repo, not roadmap promises.
M1
Memory stays organized on its own
Four background engines keep the memory store healthy: one lets stale memories fade naturally, one builds a layered table-of-contents, one previews what's safe to compress, and one extracts reusable procedures. Nothing gets deleted without your approval.
Forgetting · Layering · Compression · Procedural
M2
Search just gets the right answer
Keyword and semantic results are now blended with reciprocal-rank fusion by default on the recommended profiles, so you don't have to pick a mode upfront.
RRF on by default for profiles B / C / D
M3
Real vector search, no extra service
The native vector engine ships inside SQLite — no Pinecone, no Milvus, no separate process. Falls back gracefully if the optional library isn't installed.
sqlite-vec auto-enabled on profiles C / D
M4
No more silent overwrites
If two agents try to update the same memory at the same time, the second one gets a clear conflict back instead of clobbering the first. You decide what to do.
Optimistic CAS · returns HTTP 409 on stale writes
M5
Errors that don't leak internals
Every error response is now a clean structured message — no raw stack traces, no internal IDs, no timing leaks on review tokens. Safe to show to users, safe to log.
{ "error", "reason" } shape · timing-safe token compare
M6
Boots cleanly anywhere
Windows, macOS, Linux, and Docker all start the same way. Port already in use? It moves out of the way. Bash and PowerShell scripts behave identically. Containers no longer run as root.
SSE rate-limited · port auto-fallback · .sh + .ps1 paired
Capabilities
What you get on day one.
Each card describes something the repository already does — clone, start, and it's there.
M
Memory that survives restarts
Every memory is saved to a local file and has a clear path you can read or rename. Aliases let one memory live in multiple places, and a "recently changed" view shows you what moved.
9
One toolset for every agent
Nine tools cover everything an agent needs. The same set works in Claude Code, Codex, Gemini CLI, OpenCode, and major IDE hosts — no per-client glue.
- Read & write: read_memory, create_memory, update_memory, delete_memory, add_alias
- Search & maintain: search_memory, compact_context, rebuild_index, index_status
WG
Every write is checked & reversible
Writes pass through a guard before they land, and every change gets a snapshot. If something goes wrong, the Review page rolls it back in one click.
UI
A built-in admin panel
Open the Dashboard in your browser and you get four pages: Memory Browser, Review & Rollback, Maintenance, Observability. No separate tool to install.
A-D
Grow from laptop to production
Four ready-made profiles take you from "tiny local test" to "full hybrid search with a remote embedding API." Change one env file, switch profile, restart — that's it.
i18n
English & Chinese, end to end
Every doc has paired English and Chinese versions, and the Dashboard auto-detects your browser language. v3.9 finished UI translation parity and lets you store memories in CJK characters.
E×4
Memory keeps itself tidy
Four background engines watch over the store: they let stale notes fade, build a layered table-of-contents, suggest what's safe to compress, and extract reusable procedures. Nothing is ever deleted without your approval.
SET
Guided first-run setup
The first time you open the Dashboard, a small wizard walks you through entering an API key. If you cloned locally, it can also save common settings straight into your .env.
Architecture
Backend, MCP, dashboard, and index lifecycle stay on one path.
The public architecture is already documented in the repo. The page below links the main moving parts instead of collapsing them into a generic “AI memory platform” slogan.
Writes, reads, search, snapshots, and maintenance all live in a single backend service that stores everything in one SQLite file. Vector search runs inside that same file — no extra database to manage.
FastAPI · SQLAlchemy · sqlite-vec
MCP is the shared protocol: Claude Code, Codex, Gemini CLI, OpenCode, Cursor, Windsurf, VSCode and others all talk to Memory Palace using the same nine tools — you set it up once.
The Dashboard is just a web page — open it in any browser. If the backend disconnects, you see a banner. If one view crashes, the error boundary keeps the rest of the UI alive.
React · Vite · TailwindCSS · Framer Motion
If search quality drifts, the Maintenance page shows you the index status and lets you rebuild it. When you change profile or embedding dimension, the index detects the change and reset itself — never a silent downgrade.
Validation and docs
Everything above is checked, not promised.
The numbers come from real test runs. The "not yet claimed" list keeps us honest about what we haven't verified.
✓
v3.9 test snapshot
- Backend: 1605 tests pass in the documented v3.9 baseline. Frontend: 232 tests pass in the latest dashboard rerun.
- New tests cover write conflicts, clean error responses, rollback edge cases, search quality metrics, and the Docker one-click flow.
- The frontend builds clean on Linux, macOS, and Windows in CI.
DOC
Docs already in the repo
- Bilingual README files and a bilingual docs folder.
- Quick start, tool reference, deployment profiles, dashboard guide, evaluation, troubleshooting, security notes, technical overview — each in both languages.
- Per-release changelogs in docs/changelog/. This landing page itself deploys from docs/.
PATH
What we don't claim
- The profile benchmark tables were not re-run for v3.9, so we don't quote MRR or latency numbers here.
- Running natively on Windows or Linux (outside Docker, outside CI) still needs hands-on checks.
- OpenCode, Cursor, and Antigravity aren't claimed all-green. macOS + Docker is the most-tested combo.
Start here
Pick the way that fits you.
Three ways in: let an agent install it for you, run it yourself in five minutes, or read the tool reference first.
01
Let the agent install it for you
Want the AI to do the setup work? Start with the memory-palace-setup repo — it walks you through everything step by step.
02
Run it yourself
Local dev or Docker — the quick start has bilingual instructions. Stick with Profile B as your first try; it works without any external services.
03
Read the tool reference
Integrating a client or doing a safety review? Jump straight to the nine-tool MCP reference and the deployment-profile docs.