
For people who write code · runs locally · code never uploaded
Your AI remembers the project.
Your code never leaves the machine.
Yesterday you said pnpm. Today it still knows. The source index lives in memory; memories land in ~/.loong-recall/.
1,284
memories · 37 decisions · 12 projects
Windows / macOS / Linux · v0.11.1 · 8 AI IDEs supported · no telemetry
What it remembers
Four memory types, each with a type, a source and a time.Four memory types, each with a type, a source and a time.
Why these results
Not guessed by similarity — derived from record fields.Not guessed by similarity — derived from record fields.
Same event_id
Memories written during the same task are linked by construction.
Shared concrete entities
Both mention rate_limit.rs, so they belong to the same thread.
Same project, same time window
Decisions from the same hour and repository are read together.
Three clues: event_id · shared entities · same project and time window
Across sessions, without amnesia
The same memory is still retrieved three days later.The same memory is still retrieved three days later.
Day 1
You say: pnpm, port 3099
preference
Day 3
New session, the AI follows pnpm
recall hit
Day 7
Switch projects, old rules do not leak
project isolation
Code and memory, one store
Ask where rate limiting lives; it returns the code and the decision behind it.Ask where rate limiting lives; it returns the code and the decision behind it.
// rate_limit.rs
pub struct TokenBucket {
capacity: u32, // 100
refill_per_sec: u32,
}Rate limiting uses a token bucket at 100 req/min, placed in the middleware layer so it does not leak into business code.
Same task · shared entity rate_limit.rs · 2026-11-06
Retrieval fundamentals on LongMemEval
Numbers come from a reproducible benchmark report in the repository.Numbers come from a reproducible benchmark report in the repository.
0.900
Session Recall@10
27 / 30
0.683
Session MRR
0.192s
Mean retrieval time / instance
Methodology, limitations and all sources +
Method: LongMemEval-S public dataset, stratified sample covering 6 question types with 5 instances each, 30 in total; deep path + bge-base-zh, Top-K=10 (README.md:150).
Limitations: 30 instances is not the full 500; the stratified sample over-weights rare question types, so the numbers are not equivalent to a full weighted score; latency depends on the test machine (README.md:240–242).
Scale reference: fast path + TF-IDF over all 500 instances gives Session Recall@10 85.7% and Turn Recall@10 61.7% (README.md:178).
Regression gate: built-in benchmark 11 / 11 passed, 0 failed (v0.9.7, 5.52s, README.md:110).
Primary metric sources: README.md:154, 156, 159.
Works with any MCP-capable coding tool
The desktop app writes the configuration for you — no hand-editing JSON.The desktop app writes the configuration for you — no hand-editing JSON.
One-click configuration writing supports 8 tools; the README adds “etc.” to signal the list is not exhaustive (README.md:374–383).
Code and memory do not leave the machine on their own
The only outbound case: if you configure LLM query translation, only the query text is sent.The only outbound case: if you configure LLM query translation, only the query text is sent.
- Local storage
- Memories in ~/.loong-recall/; embedding model weights by default in ~/.loong-recall/models/.
- No reporting
- No telemetry, no tracking.
- Index not persisted
- The source index stays in memory.
- Exportable
- The desktop settings page provides backup and export.
Using DeepSeek Harness?
Offload conclusions into LRC so the context stops growing.Offload conclusions into LRC so the context stops growing.
- What it does
- Every 8 turns / 6000 characters it distills decisions, constraints and conclusions, deduplicates them, writes them into LRC, and leaves a one-line status snapshot.
- Three acceptance checks
- Incremental cursor < 50%; three repeated migrate calls do not grow the store; after three offloads the total is ≤ 1.2× the first run with ≥ 80% deduplication.
- Boundary
- It does not promise “never compacts”. Compaction is still produced by DSH itself; the plugin only reduces how often it triggers.
Two ways to install
Commands and notes come from the repository README.Commands and notes come from the repository README.
Option 1 · recommended
Download the desktop app
Download the lrc-desktop-* package and double-click to install. The desktop app detects your AI tools, writes the MCP configuration and rule files, and you restart the IDE.
- Windows 10 / 11 (x64)Download .exe
- macOS (Apple Silicon)Download .dmg
- Linux (x86_64)Download .deb
Version v0.11.1. Checksums are in the .sha256 file in the same directory, and also on GitHub Releases.
Option 2 · license boundary
Why only installers
The shell, frontend and storage modules are public under Apache 2.0. The core retrieval engine (encoding, mirror trapezoid, DaoTi regulation, retrieval fusion) is not public and ships only as a compiled artifact inside the installer. The public repository references the private core as a submodule, so an unauthorized clone cannot build.
- Shell / frontend / storage: Apache 2.0, public
- Core retrieval engine: DaoTi Research License, not public
- Distribution: GitHub Releases and this site only
Source: README.md, “License and source boundary”.
Questions you may have
Tool integration, data location and network boundaries.Tool integration, data location and network boundaries.
Which AI tools are supported?+
Eight are built in, including Trae, Cursor, VS Code and Windsurf. Any tool that speaks MCP can connect.
Where is data stored?+
Memories in ~/.loong-recall/; embedding model weights by default in ~/.loong-recall/models/.
Does it need a network connection?+
Basic retrieval does not. Only semantic search downloads a model on demand, and query text is sent only if you configure LLM query translation.
Can I export my data?+
Yes. The desktop settings page provides backup and export, and snapshot restore since v0.10.2.
Why are there no associations right after install?+
Associations are driven by raw material. An empty library returns an honest empty state rather than forcing links; in a real library the event_id coverage is only 1.23%.
Can it be hosted remotely?+
Not recommended. It depends on a local index, a local memory store and a local model.
Is Loong Recall open source?+
No. The desktop shell and web frontend are Apache 2.0; the core retrieval engine is closed and shipped only as a compiled artifact, so only installers are provided, not source builds.
Does it compress context?+
No. LRC does not compress; it distills conclusions into the local memory store before compression, recalls them on demand afterwards, and leaves an [archived] marker so the model knows where the details are. It does not promise that compression never happens.
How should the benchmark numbers be read?+
Three separate settings: 0.900 is LongMemEval-S 30 instances (deep path); 94.47% is the v0.10.4 fast path over 470 valid instances; 85.74% is the v0.5.6 historical baseline. They cannot be added together or substituted for one another.
How do the DSH plugin and the desktop app relate?+
The plugin is the entry point; the desktop app is the memory backend. dsh-plugin-lrc 0.2.1 (MIT) distills conclusions from DSH sessions into LRC without changing DSH's native compression.
Give your AI cross-session local memory
Fully local, zero cloud dependency.