Vector search & agent memory — what's accelerating
"Memory" has become the hottest word in agent tooling — but it spans everything from a JSON file of session notes to a distributed vector database. The repos below are the fastest-climbing projects that actually store and retrieve agent state, sorted from purpose-built memory layers down to the vector databases underneath them.
Top mover
A free, open-source AI memory system built on ChromaDB and exposed over MCP, leading on public memory benchmarks. The benchmark claim is what separates it from the crowd of memory wrappers — it competes on recall accuracy, not just on having a tidy API, and the MCP surface means any agent can plug in without custom glue.
---
The memory & vector stack
Persistent context across sessions for any agent: it captures what the agent does, compresses it with an LLM, and re-injects the relevant parts next session. The compression step is the real work — raw transcripts overflow the context window fast, so the value is in what it throws away.
Persistent memory aimed specifically at coding agents, positioned on real-world benchmark results. Narrower scope than a general memory layer — it's tuned for the code-context case (files touched, decisions made) rather than open-ended chat history.
A "memory upgrade" for coding agents, shipped as a Go binary. The language choice matters for adoption: a single dependency-free binary is far easier to drop into a CI or local toolchain than a Python package with its own environment.
A universal memory layer for agents and the most established name in this list. Accelbrief has cooled as it matures, but it remains the default many teams reach for — broad integrations and a stable API are worth more than raw growth rate once you're in production.
The vector database under many of the layers above — high-performance ANN search in Rust, with hybrid (vector + keyword) retrieval. Low stars/day because it's infrastructure that already won its niche; you adopt it for stability, not novelty. The alternative path is pingcap/tidb (⭐40,137 · ↑10.2/day), which now bolts native vector search onto a transactional SQL database — one store for both your app data and your embeddings instead of running a separate vector service.
---
Context: high velocity, wrong shelf
The two highest-velocity repos tagged into this bucket aren't memory infrastructure — track them as signals, not building blocks: - affaan-m/ECC (⭐208,769 · ↑1,501.9/day) — tagged with memory, but it's an agent-harness performance/optimization bundle (skills, instincts, security), not a memory store you'd integrate. Trend signal, not infrastructure. - 666ghj/MiroFish (⭐64,955 · ↑338.3/day) — a swarm-intelligence prediction engine that happens to carry agent-memory and knowledge-graph tags; the actual product is forecasting, not retrieval. - zhayujie/CowAgent (⭐45,084 · ↑32.2/day) — a full agent assistant with memory as one feature, not a memory layer you'd reuse elsewhere.
The pattern: "memory" is now a marketing tag attached to anything agentic. The genuinely reusable layers — MemPalace, claude-mem, mem0 — move slower than the all-in-one harnesses but are what you actually build on.
---
How this was made
Live GitHub pull, bucketed by theme, verified not-archived and pushed recently, ranked by stars/day, curated for substance. Counts pulled at publish — they move daily.
Accelbrief · catch acceleration, not stars · all editions