palace-rs

A local-first memory retrieval engine for coding agents, implemented in Rust.

This project stores verbatim project and conversation memory, embeds it locally, and retrieves source-grounded context through MCP. It is built for coding agents that need to remember decisions, prior fixes, commands, project conventions, and user preferences across sessions without running a separate vector database.

What It Does

• Stores project files and conversation turns in a local SQLite database.
• Generates local embeddings with ONNX Runtime and all-MiniLM-L6-v2.
• Retrieves memories with hybrid semantic/BM25 search plus coding-agent intent boosts.
• Tags preference-shaped drawers and runs a dedicated preference recall pass for fuzzy "what do I prefer?" and convention questions.
• Stores preference spans with optional secondary embeddings and exposes a preference_match score for preference-shaped queries.
• Classifies search intent (preference, decision, how_to, definition, temporal, unknown) and can optionally rerank top results with a local interaction reranker.
• Sanitizes agent-generated query dumps before retrieval.
• Returns source-grounded results with score provenance and nearby source context.
• Warms up agents from recent diary entries with project, topic, timestamp, tags, and compact session text.
• Provides a knowledge graph for temporal entity relationships.
• Measures real-world usefulness through palace gain precision metrics and optional folded feedback on the existing palace_gain MCP tool.
• Exposes MCP tools for assistants that support Model Context Protocol.
• Offers a small Rust library API for embedding memory into other services.

Agent Memory Reliability

Version 0.6.0 focuses on the retrieval cases that matter most during coding work: preferences, project conventions, recent session continuity, source-grounded answers, and measurable usefulness in real agent sessions. Drawers that look like user preferences or conventions are tagged in metadata during writes and updates, record the matched preference span, and can store a secondary preference embedding. Preference-shaped queries receive a dedicated preference_match score alongside hybrid semantic/BM25 search.

MCP search responses expose score provenance (combined, cosine, bm25, and coding_boost, preference_match, optional rerank_score, and intent) plus adjacent source context so agents can cite why a memory was returned. Diary tools provide warm-start context for recent sessions, including project path, topic, timestamp, session ID, tags, and compact text.

Library consumers can use Palace::search_with_provenance when they need the same structured score details that MCP tools return.

Storage

Collapses Python's dual-store (ChromaDB + SQLite) into one file at ~/.palace/palace.db:

Embeddings are stored as f32 vectors from all-MiniLM-L6-v2. Search uses local cosine similarity over the stored vectors.

Benchmarks

Coding-Agent Memory Eval

The repository includes a focused eval fixture for practical coding-agent memory questions. It stores realistic memories about project decisions, prior failures, commands, conventions, user preferences, and current direction, then asks 40 questions such as:

• why did we choose bundled sqlite?
• how did we fix the migration test failure last time?
• what clippy command should I run?
• what is the project convention for search results?
• what changed in the current product direction?

Run it with:

cargo test --test coding_agent_eval -- --nocapture

The test reports recall@1 and recall@5 and fails if retrieval drops below the stable threshold. This is the product-shaped proof: not broad memory theater, but whether a coding agent can recover the right project context when it matters.

LongMemEval

Retrieval recall on the LongMemEval s_cleaned split — 500 questions over conversational haystacks of ~50 sessions / ~115k tokens each (30 abstention questions are filtered out per the standard convention, leaving 470 evaluated).

The recipe behind the numbers below:

• Granularity: one drawer per session.
• Indexed content: the full session — both user and assistant turns are stored and embedded together. No user-turn filtering, no summarization, no LLM extraction.
• Embedder: all-MiniLM-L6-v2 (384-dim, ONNX), 512-token cap, run locally — no API calls.
• Retrieval: hybrid baseline — BM25 (k1=1.5, b=0.75, weight 0.35) fused with cosine similarity (weight 0.65), top-K = 10. These reported LongMemEval numbers used pure score fusion, before the coding-agent intent boosts used by current project-memory search.
• No LLM at any stage: no extraction, no rerank, no answer generation. The recall numbers measure the retriever in isolation.
• Metric: recall_any@K at session granularity — does any gold session appear in the top-K results?
• Hardware: Apple M1 Pro, 10 cores (8P + 2E), 32 GB RAM.

Per-question-type on s_cleaned:

Reading the numbers

• oracle is a sanity check, not a real result. That split hands the retriever only the sessions known to contain the answer, so perfect recall just confirms the pipeline is wired up correctly.
• s_cleaned is the real test. ~50 sessions / ~115k tokens of conversational haystack per question, no hints. R@5 = 0.981 means that for 461 of 470 evaluated questions, a gold session appears somewhere in the top 5 retrieved.
• R@1 → R@5 → R@10 tells you where the failures cluster. The jump from 0.889 to 0.981 means most "misses" at top-1 are near-misses — the right session is usually rank 2–5, displaced by a lexically similar distractor. The further jump to 0.991 at top-10 means only ~9 questions out of 470 fall outside the top-10 entirely; those are the genuinely hard cases.
• Per-question-type breakdown is where the model's blind spots show.
  • single-session-assistant, single-session-user, knowledge-update: ≥0.94 at R@1, ≈1.0 at R@5. The retriever handles direct questions where the answer is stated verbatim in one session.
  • multi-session and temporal-reasoning: strong at R@5 (~0.98) but lower at R@1 (~0.83–0.91). Multiple sessions are relevant and the "best" one is a judgement call — top-1 ranking among near-equivalents is genuinely ambiguous.
  • single-session-preference: the visible weak spot at 0.633 / 0.867 / 0.933. Preference questions ("what's my favorite X") are answered by sentences like "I like…" / "I prefer…" that don't share keywords with the question. Pure BM25 + frozen MiniLM has no signal for preference-shaped sentences specifically; closing this gap would require either an LLM-extracted preference index or a hand-rolled pattern booster.
• What's deliberately not in these LongMemEval numbers. No LLM at any stage — no extraction during ingest, no query rewriting, no rerank, no answer generation. No per-dataset hyperparameter tuning. No GPU. The result is the baseline retriever in isolation, on a single CPU, with fixed defaults.

Installation

Homebrew (macOS Apple Silicon / Linux)

brew tap AncientiCe/palace
brew install palace

# Configure MCP servers
palace install --all

Note: macOS Intel is not supported due to ONNX Runtime unavailability. Apple Silicon and Linux x86_64 are fully supported.

Install Script (macOS / Linux / Windows)

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/AncientiCe/palace-rs/main/scripts/install.sh | sh

Windows:

irm https://raw.githubusercontent.com/AncientiCe/palace-rs/main/scripts/install.ps1 | iex

The installer downloads the matching GitHub Release binary, verifies its SHA-256 checksum, installs it locally, and registers the MCP server with Cursor, Codex, and Claude Code.

Development Install

cargo install --path .
palace install

The first time you run mine, the embedding model is downloaded automatically from HuggingFace and cached.

Upgrading from mempalace (≤ 0.1.9)? See Migrating from mempalace to palace. The legacy mempalace shim binary and MEMPALACE_* env vars were removed in 0.3.0 — install 0.2.x first if you need the automated migration path.

Quick Start

cargo install --path .       # development install; release installers do this for you
palace install               # configures Cursor + Codex + Claude Code
palace doctor                # verifies MCP config, rules, binary, and drawer count
palace seed-adoption-facts   # seed KG facts that make agent recall measurable
palace init ~/my-project     # detect rooms and write palace.yaml
palace mine ~/my-project     # populate the palace

Then restart your agent app so it reloads MCP configuration. Search manually with palace search "how did we decide on the database schema" or let your agent call the MCP tools when its installed rule tells it to consult memory.

CLI Reference

mine flags

palace mine ~/my-project \
  --wing my_project          # Override wing name
  --limit 100                # Cap at 100 files
  --dry-run                  # Preview without storing
  --no-gitignore             # Ignore .gitignore rules
  --include vendor,third_party  # Force-include these paths

mine-convos flags

palace mine-convos ~/Desktop/transcripts \
  --wing claude_sessions \
  --mode exchange   # or: general (decisions/milestones/emotions)
  --limit 50
  --dry-run

split flags

palace split \
  --source ~/Desktop/transcripts \
  --min-sessions 2 \
  --dry-run

gain

palace gain summarizes automatic MCP usage by Cursor, Codex, Claude Code, or any other MCP client. It records local tool-call metadata in palace.db and estimates value from retrieval hits, duplicate skips, KG facts, diary recalls, repeat questions, and latency.

palace gain
palace gain --project my_project --since 7d
palace gain --history
palace gain --json
palace gain --record <query_id> <drawer_id> useful

Example output:

palace gain - last 30d (palace_rs)
  Tool calls         : 412   (sessions: 27)
  Hit rate           : 88%   (search hits 142/162)
  Precision@1        : 92%
  Precision@5        : 95%
  Tokens saved (est) : ~78,400
  Re-index skipped   : 31    (duplicate drawers avoided)
  KG facts recalled  : 56
  Diary recalls      : 8
  Repeat Qs avoided  : 19
  p95 latency        : 41 ms
  Tool latency       : palace_search(p50 18 ms, p95 41 ms)
  Top wings          : palace_rs(120), checkout(40)

Set PALACE_GAIN_DISABLED=1 to disable usage recording.

palace_gain also accepts an optional record payload for MCP callers that want to file explicit usefulness feedback without learning a new tool:

{"record": {"query_id": "query_abc", "drawer_id": "drawer_xyz", "verdict": "useful"}}

MCP Setup

palace install is the normal setup command for the four supported local agent clients: Cursor, Codex, Claude Code, and Claude Desktop. It writes both:

• an MCP server entry that starts palace mcp
• a small rule that tells the agent when to call palace_status, palace_search, palace_preference_search, palace_kg_query, and palace_diary_write

palace install

What gets written by default:

Existing 0.1.x installs that registered the server as mempalace are migrated to palace automatically the next time you run palace install.

Install for one client:

palace install --client cursor
palace install --client codex
palace install --client claude

Install project-scoped rules instead of global rules:

palace install --scope project --path /path/to/project

For project scope, Cursor also gets a project-local MCP config at <project>/.cursor/mcp.json. Codex and Claude Code keep MCP config in their user-level config files, while their rules go into <project>/AGENTS.md and <project>/CLAUDE.md.

Skip rule files if you only want MCP wiring:

palace install --no-rule

Inspect the current setup:

palace doctor

The installed rule is memory-first for remembered context: decisions, prior fixes, conventions, preferences, prior commands, session history, and "what happened last time?" should use Palace before grep or code search. Grep remains the right first tool for current symbols, exact definitions, exact files, and implementation details that may have changed since the project was mined. It also tells agents to warm-start with palace_session_context, search diaries with palace_diary_search before continuing old work, use KG tools for durable facts, and write palace_diary_write after substantive work.

Remote mode (shared palace-server)

By default palace mcp serves the local palace. Point it at a shared remote Palace Server instead — so a whole team shares one memory backend in their own infrastructure — without changing any client's stdio registration. In remote mode palace mcp becomes a transparent stdio→HTTP bridge that forwards each request to the server's /mcp endpoint with a Bearer API key. Palace Server is the commercial, self-hosted team edition — licenses, docs, and deployment guides live at palacememory.com.

# Store the endpoint and ps_… API key (prompts for the key if --api-key is omitted)
palace remote set --endpoint https://palace.yourco.com

# Switch the MCP server to the remote palace-server, then verify
palace remote on
palace remote test          # runs the MCP handshake; reports tool count

# Back to the local palace at any time
palace remote off           # (or: palace local)

Inspect the current wiring with palace remote status (prints the MCP mode, the normalised /mcp endpoint, and a masked API key). Remote settings are read from the PALACE_MCP_MODE, PALACE_REMOTE_ENDPOINT, and PALACE_API_KEY environment variables, falling back to the mcp_mode, remote_endpoint, and remote_api_key keys in ~/.palace/config.json (written with owner-only 0600 permissions). The endpoint accepts a bare host, a base URL, or a full /mcp URL.

Automatic memory hooks

palace install registers user-scope hooks for every client that supports them, so memory use is automatic in every project without per-project rule edits. The three hooks behave the same everywhere:

• session start — injects the protocol (Cursor also exports PALACE_SESSION_ID).
• post tool use — auto-recalls relevant memory while the agent investigates, so a prior agent's decisions surface even before the agent thinks to search.
• stop — if the session engaged Palace but recorded nothing, it asks the agent to palace_diary_write its investigation and palace_kg_add durable decisions before finishing. It nudges at most once.

Claude Code and Codex share a "Claude-style" output dialect (hookSpecificOutput.additionalContext for context, decision: "block" + reason to keep the agent working until it saves); Cursor uses its own additional_context / followup_message keys. The runner that produces these is palace hook <event> --client <cursor|claude|codex>.

Cross-agent continuity: palace_diary_search accepts all_agents: true (and an optional project_path) to recall investigations recorded by any agent, and palace_session_context falls back to another agent's recent work for the project when you have none of your own. Durable decisions belong in the knowledge graph (palace_kg_add / palace_kg_invalidate), which dedupes facts and tracks changes over time, so re-recalled decisions never duplicate.

Seed durable KG facts for adoption tracking:

palace seed-adoption-facts --project my_project

The seed is idempotent and records the four supported clients, the memory-first protocol, routing rules, user preference for memory-aware agents, and standard quality gates. Agents can then recall those facts with palace_kg_query.

Remove palace config:

palace uninstall
palace uninstall --client cursor

Cursor

After palace install --client cursor, restart Cursor or reload the window. Settings -> MCP should show palace as an enabled stdio server.

Manual Cursor config shape:

{
  "mcpServers": {
    "palace": {
      "command": "palace",
      "args": ["mcp"]
    }
  }
}

The rule is installed as .cursor/rules/palace.mdc with alwaysApply: true.

Codex

After palace install --client codex, restart Codex so it reloads ~/.codex/config.toml.

Manual Codex config shape:

[mcp_servers.palace]
command = "palace"
args = ["mcp"]

The rule is installed as a managed palace block in ~/.codex/AGENTS.md (or <project>/AGENTS.md with --scope project). Existing content is preserved.

Claude Code

After palace install --client claude, restart Claude Code so it reloads ~/.claude/mcp_servers.json.

Manual Claude JSON shape is the same as Cursor's mcpServers object above. You can also use Claude Code's own MCP command:

claude mcp remove palace
claude mcp add palace -- palace mcp

The rule is installed as a managed palace block in ~/.claude/CLAUDE.md (or <project>/CLAUDE.md with --scope project). Existing content is preserved.

MCP Tools

The server exposes tools for status, taxonomy, search, drawer CRUD, knowledge graph operations, graph tunnels, hook acknowledgements, and agent diaries:

Migrating from mempalace to palace

The 0.2.0 release renamed the project from mempalace to palace. The 0.2.x line kept the old names working with deprecation warnings; they were removed in 0.3.0. On current versions, migrate via a 0.2.x release first or rename manually (~/.mempalace → ~/.palace, mempalace.yaml → palace.yaml).

One-step migration:

cargo install --path .
palace install

palace install rewrites existing MCP client configs (Cursor, Codex, Claude Code) and rule files, replacing legacy mempalace entries with palace entries. ~/.mempalace is moved to ~/.palace on first run when the legacy directory exists and the new one does not.

Migration from Python

The Rust version uses a new single-file database (palace.db). Your existing ChromaDB data cannot be migrated automatically.

Steps:

# 1. Re-mine your projects
palace init ~/my-project && palace mine ~/my-project

# 2. Re-index conversations
palace mine-convos ~/Desktop/transcripts

# 3. Verify
palace status

Your identity.txt, people_map.json, and known_names.json in ~/.palace/ (migrated from ~/.mempalace/ if present) are compatible and will be read automatically.

Test on a Project

palace init /path/to/project
palace mine /path/to/project
palace status

Restart Cursor, Codex, or Claude Code, then ask the agent a project question that should use memory, for example: "Search the palace for how this project handles database migrations." The agent should call palace_search through MCP instead of re-indexing the repository from scratch.

Configuration

~/.palace/config.json is read on startup. Environment variables take highest priority:

palace.yaml (per-project)

Created by palace init. Example:

wing: my_project
rooms:
  - name: backend
    description: Server and API code
    keywords: [api, server, routes, models]
  - name: frontend
    description: UI components
    keywords: [ui, components, pages, views]
  - name: general
    description: Everything else
    keywords: []

Memory Stack

palace wake-up prints L0 + L1. The AI uses MCP tools for L2/L3.

Development

cargo build
cargo test
cargo clippy

Tests use in-memory SQLite — no palace.db needed. The embedding model is not loaded in tests that don't require it.

Hooks Compatibility

Shell hooks that previously called python -m mempalace.mcp_server or mempalace mcp can now call palace mcp. Update the binary path in your hooks:

# Before (Python)
exec python -m mempalace.mcp_server

# Before (Rust 0.1.x)
exec mempalace mcp

# After (Rust 0.2.x+)
exec palace mcp