Observed Signal · Mar 26, 2026 · Technical Release · Source: DEV Community · Impact: 2/5 · Sentiment: Positive
Vector Memory for Claude Code
An engineer replaced large, static CLAUDE.md files with a server-side vector memory system for Claude Code. They built an open-source MCP server, claude-memory-mcp, that stores structured memories in Supabase + pgvector and uses OpenAI's text-embedding-3-small for semantic retrieval. The system exposes MCP tools (remember, recall, forget, project_status) over HTTP (port 3101), enforces structured memory types (eight categories), imposes a 2,000-token recall cap, and implements soft-deletes and TTLs. CLAUDE.md remains for stable rules while vector memory holds evolving decisions, bug fixes, and patterns; Claude is instructed to call the MCP tools during sessions. The author reports low estimated monthly embedding costs (~$0.50) for typical consulting volumes and provides a 15-minute setup path using Docker, Supabase, and an OpenAI key.
Provides a reproducible architecture pattern for LLM session memory (server-side vector memory + MCP) that can reduce context bloat and enable cross-project semantic recall; useful for teams adopting LLMs but not a major platform policy or industry-shifting announcement.
Track OpenAI Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Author created claude-memory-mcp, an open-source MCP server for Claude Code.
- Storage backend: Supabase with pgvector for vector search.
- Embeddings: OpenAI text-embedding-3-small (1536 dimensions).
- Exposed MCP tools: remember, recall, forget, project_status available over HTTP (port 3101).
- Memories are typed into eight structured categories (decision, bug_fix, pattern, context, blocker, learning, convention, dependency) and recall responses are capped at 2,000 tokens.
Connected Companies & Entities
3 Entities mappedOntology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Anthropic ships memory into Claude Code
A developer describes building an MCP-based persistent memory layer for AI coding assistants called cachly and how Anthropic's announcement that Claude Code gained a memory feature prompted re-evaluation of that work. The author explains a reproducible test for measuring whether an assistant truly persists facts between sessions, shares benchmark results from a personal corpus, and summarizes community feedback that changed the project's metrics and reliability checks. The post notes cachly offers a free tier hosted in the EU and links to the project's site.
Lossless AI Memory Tool for Claude Code
Longhand is an open-source tool that captures Claude Code session logs verbatim, indexes them locally, and exposes deterministic recall without LLM summarization. Instead of relying on larger context windows, Longhand ingests JSONL session files and stores structured events in SQLite and semantic vectors in ChromaDB. It runs an auto-ingest hook on session end, supports one-time backfill, and exposes recall via an MCP server with ~17 Claude tools (recall, search_in_context, get_session_timeline, replay_file, etc.). The project is published on PyPI (longhand), registered in the MCP Registry, MIT-licensed, Python 3.10+, and tested against 107 sessions (53,668 events). Reported characteristics: ~126ms semantic recall, ~200–400MB typical storage (up to ~1GB heavy users), zero network/API calls, offline operation, and 170 unit tests with security audit results showing no critical findings.
MySQL Journal Fixes Claude's Session Amnesia
A developer describes a lightweight approach to give Anthropic's Claude persistent session context without relying on heavyweight vector stores. The solution stores structured 'journal' entries in a flat MySQL table and exposes them via three MCP (Model Context Protocol) tools (journal_get, journal_post, journal_get_all). Claude reads the journal at session start via mcp-remote, restoring user instructions and recent context each session. The post includes the dev_journal table schema, a deterministic RK key generator using MD5, MCP server setup notes (MapMcp needs a leading slash; options.Stateless = true), nginx proxy settings, deployment options (ngrok, VPS, Raspberry Pi), and links to source code on GitHub. The author recommends this simpler pattern for remembering coding standards and project context instead of vector DBs or semantic search pipelines.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.