Importer lookup
Approximately 45,000 input tokens reduced to about 350 through structural lookup.
Agent memory · MCP · Local first
Project Memory
A durable knowledge layer for coding agents that retrieves precise code structure, decisions, research, and reusable scripts instead of repeatedly loading entire repositories.
85–99%fewer input tokens in documented targeted query comparisons
90.6%embedding storage reduction: 1,536 to 144 bytes per vector
88.9%top-3 recall reported for the TurboQuant configuration
<1msdequantization time per vector in documented measurements
System architecture
Memory is a retrieval system, not a transcript dump.
Thin proactive hooks capture durable facts while 11 MCP tools expose knowledge on demand. Structural code search and persisted project history are ranked through hybrid retrieval, compressed, and returned as targeted context.
Measured economics
Spend context where reasoning needs it.
Documented comparisons separate input-token reduction, vector-storage compression, and optional local-routing economics rather than blending them into one headline.
Function understanding
Approximately 11,000 input tokens reduced to about 700 with targeted code context.
Session start
Approximately 50,000 input tokens reduced to about 100 for focused start-of-session context.
Decision recall
Approximately 2,500 input tokens reduced to about 200 using durable decision memory.
Script reuse
Approximately 2,000 input tokens reduced to about 300 by retrieving proven automation.
Optional local routing
The documented Ollama router can shift 35–55% of typical session cloud spend locally.
Technology foundation
Local-first retrieval, built from composable parts.
Project Memory combines the Model Context Protocol with structural parsing, embedded search, vector inference, compressed embeddings, and optional local model routing.
Node.jsServer, daemon, and tooling runtime
MCP11 on-demand agent tools
SQLite FTS5Structural graph and lexical search
ONNXLocal semantic embeddings
Tree-sitterLanguage-aware code structure
OllamaOptional local-first AI routing
Solution fit
Retrieval beats reloading the repository.
The common answer to agent memory is a bigger window and a bigger paste. Project Memory treats memory as a retrieval problem instead: durable facts are captured once, code is indexed structurally, and each question is answered with the smallest sufficient slice.
Agents that start every session from zero
Decisions, research, and proven scripts persist across sessions and across projects. A focused start-of-session context is documented at roughly 100 input tokens where a cold reload cost about 50,000.
Questions that should not cost a whole repository
A Tree-sitter structural graph, ranked through hybrid lexical and semantic retrieval, returns the specific relationship. An importer lookup drops from roughly 45,000 input tokens to about 350.
Knowledge that has to stay on the machine
SQLite and ONNX keep the primary stores local, inspectable, and tool-addressable, and the optional Ollama router can move 35–55% of typical session spend off the cloud entirely.
| Capability | Typical market approach | Project Memory |
|---|---|---|
| What persists | A rolling transcript window; anything beyond it is gone. | Durable decisions, research, code structure, and reusable scripts held across sessions and across projects. |
| How code reaches the model | Whole files — often whole repositories — pasted into context. | A Tree-sitter structural graph returns the specific relationships instead of the file. |
| Retrieval method | Embedding similarity alone, which misses exact identifiers. | Hybrid retrieval: SQLite FTS5 lexical precision ranked together with ONNX semantic search. |
| Cost of storing memory | Full 1,536-byte float vectors for every embedding. | TurboQuant compression to 144 bytes — 90.6% smaller — at 88.9% documented top-3 recall and sub-millisecond dequantization. |
| Where the data lives | A vendor-hosted memory service outside the customer’s control. | Local-first: the primary stores stay on the machine, inspectable and portable. |
| Agent compatibility | A proprietary plugin written per assistant. | Eleven Model Context Protocol tools, usable by any MCP-capable agent. |
Comparison note: the market column describes the default behaviour of transcript-window and vendor-hosted memory approaches rather than any named product. Token, storage, recall, and latency figures come from separate documented project measurements of different optimization dimensions and should not be added together.
Related systems
Metric note: token, storage, recall, latency, and local-routing figures come from separate documented project measurements. They describe different optimization dimensions and should not be added together.