Purpose
0CompactMem (also called vMem in the repository) provides persistent memory for Claude Code and various LLM agents, solving the problem of agents losing decisions, constraints, and historical understanding after context compaction. It treats the context window as a hot working set and places persistent knowledge outside the window, swapping it in on demand when needed, achieving "operating-system-style" context continuity, so that memory can be retained across sessions, across compactions, and across agents.
Main capabilities
- Persistent multi-agent memory, stored in a single-file SQLite (WAL mode) database, with built-in FTS5 full-text indexing.
- Draws on operating system primitives: demand paging, kswapd-style eviction, mlock-style pinning, capacity-aware reclamation.
- BM25 retrieval and importance scoring.
- Multi-agent sharing: provides a cross-agent notify bus and IPC.
- MCP native support.
- No manual memory management required at agent runtime.
- No GPU, no external API; everything runs locally.
Usage
Officially recommends a one-line command installation via Claude Code; manual installation is also supported. After installation, the entire pipeline runs inside Claude Code hooks, which sit at Claude Code's system call boundaries (SessionStart, UserPromptSubmit, Stop, PostToolUse) and call into the storage layer. For detailed hook configuration, daemon management, and troubleshooting, see docs/SETUP.md; for architecture, see docs/ARCHITECTURE.md; for design philosophy, see docs/DESIGN_PHILOSOPHY.md. The text mentions pip install as a clue to the usage method, but the specific installation command content is not yet provided.
Architecture and subsystems
- Hook layer: located at Claude Code's system call boundaries.
- Storage layer: a single SQLite file, unified VFS interface (memory_os.store.api / memory_os.store.vfs / memory_os.store.criu).
- Daemon and IPC: persistent retrieval daemon (Unix socket), asynchronous extractor pool (kworker-style), cross-agent notify bus.
Applicable scenarios
- Developers who frequently encounter"context compacted" interruptions when using Claude Code.
- Teams that need multiple agents to share learned knowledge and avoid each starting from scratch.
- Scenarios where memory capability should run locally without relying on external APIs or GPUs.
- Workflows that need to pin key constraints so they are not evicted.
Validation and performance
The repository provides paper-style evaluation descriptions against mem0 / Letta / Zep (arXiv preprint, referred to in the materials as "formal evaluation"), as well as a performance overview; specific figures are not yet provided. Implemented capabilities cover per-session database isolation, concurrent write safety, cross-agent IPC delivery, extractor pool queue semantics, CRIU checkpoint validation, and target progress idempotency, with a cumulative 1,051+ tuning iterations and eight rounds of major capability building.
Limitations
- The materials do not provide pricing, cost, or commercial licensing information.
- The materials do not provide specific supported operating systems, Python versions, or hardware requirements.
- Roadmap items such as distributed vMem, adaptive watermarks, and per-block embedding routing are still planned; the materials do not state release dates.
- Specific metrics for the evaluation data are not yet provided.
