Skip to content

Comparison with Alternatives

memsearch is both a CLI engine and a set of native plugins for four coding CLIs, so we compare it against projects along that whole spectrum, plus Claude Code's built-in memory as a baseline: Claude Code native memory, claude-mem, qmd, MemPalace, mem0, Letta / MemGPT.

Verified against each project's README / official docs. The space moves fast — open an issue if anything looks stale.

At a glance

memsearch Claude Code native claude-mem qmd MemPalace mem0 Letta
Shape Engine + 4 native CLI plugins Built-in (Claude Code only) Plugin (Claude Code / Gemini CLI / OpenClaw) Engine + MCP + Claude Code plugin Claude Code plugin + MCP Library + native plugins + MCP Agent runtime (own CLI: Letta Code)
Source of truth Plain .md Plain .md (CLAUDE.md + auto-memory) SQLite + ChromaDB Plain .md ChromaDB Vector DB (+ optional graph) Postgres / git-backed MemFS (Letta Code)
Write Append-only User edits CLAUDE.md; auto-memory appended by Claude LLM-compressed transcripts — (read-only) Raw transcripts LLM-extracted facts, LLM add/update/delete Agent self-edits via tools
Search Dense + BM25 + RRF None — whole file loaded every session Chroma vector + FTS5 BM25 + dense + LLM rerank Dense Dense (+ optional rerank, + optional graph) Dense archival
Local default ONNX bge-m3, no key N/A (no search) Chroma default Local GGUF Local Llama + Chroma Needs LLM API on every write Configurable
Scale Milvus Lite → Server → Zilliz Cloud (same API) Bounded by context window Single machine Single machine Single machine Pluggable vector DB Postgres / pgvector

Feature matrix

memsearch Claude Code native claude-mem qmd MemPalace mem0 Letta
Platform plugins
Claude Code ✅ ✅ built-in ✅ ✅ ✅ ✅ ❌
OpenClaw ✅ ❌ ✅ ❌ ❌ ❌ ❌
OpenCode ✅ ❌ ❌ ❌ ❌ ❌ ❌
Codex CLI ✅ ❌ ❌ ❌ ❌ ✅ ❌
Cursor ❌ ❌ ❌ ❌ ✅ ✅ ❌
Gemini CLI ❌ ❌ ✅ ❌ ✅ ❌ ❌
Generic MCP — ❌ ❌ ✅ ✅ ✅ ❌
Storage
Markdown as source of truth ✅ ✅ ❌ ✅ ❌ ❌ ✅ (MemFS)
Git-diffable memory files ✅ ✅ ❌ ✅ ❌ ❌ ✅ (MemFS)
Search & retrieval
On-demand retrieval (not full-file reload) ✅ ❌ ✅ ✅ ✅ ✅ ✅
Hybrid BM25 + dense ✅ ❌ ✅ ✅ ❌ ❌ ❌
RRF fusion inside the vector DB ✅ ❌ ❌ ❌ ❌ ❌ ❌
Pluggable embedding providers ✅ (8: openai / google / voyage / jina / mistral / ollama / local / onnx) — ❌ ✅ ❌ ✅ ✅
Optional cross-encoder reranker ✅ ❌ ❌ ✅ (LLM rerank) ❌ ✅ ❌
Progressive disclosure: search → expand → transcript ✅ ❌ ❌ ❌ ❌ ❌ ❌
Forked-subagent recall (isolated context) ✅ ❌ ❌ ❌ ❌ ❌ ❌
Writes
No external API key by default ✅ ✅ ✅ ✅ ✅ ❌ ❌
Scale
Local → self-hosted → managed, one API ✅ ❌ ❌ ❌ ❌ ❌ ❌

Where memsearch is different

  • Covers Claude Code + OpenClaw + OpenCode + Codex CLI in one project. No other entry covers all four.
  • Retrieves on demand instead of stuffing the whole file into every session like Claude Code's built-in memory.
  • Markdown + Milvus, not an opaque DB. qmd and Letta's MemFS share the markdown-canonical approach; claude-mem / MemPalace / mem0 keep state in a DB.
  • Append-only writes, no LLM curation on the write path. mem0 and Letta's traditional memory depend on LLM write-time curation (powerful but can silently mutate past writes).
  • Hybrid dense + BM25 fused via RRF inside Milvus. qmd and claude-mem are also hybrid; mem0 / MemPalace / Letta archival are dense-only.
  • Scale path: Lite → Server → Cloud, one API. Others are single-machine or require wiring your own backend.
  • Context isolation via forked subagents on Claude Code — recall runs in its own context window.

When another project fits better

  • Only use Claude Code and memory is tiny / project-instruction-like → built-in CLAUDE.md is fine.
  • Generic LLM app, not a coding CLI → mem0.
  • Want the LLM to actively curate memory → Letta or mem0.
  • Want a full agent runtime or MemFS → Letta.
  • Cursor / ChatGPT / Gemini CLI users → mem0, MemPalace, or claude-mem.
  • Just need a local markdown search engine → qmd.

References