Cache-Pot – Sumit Kumar

Cache-Pot July 2026

AI Infra

Redis

Vector Search

A Redis-compatible memory engine for AI agents — one server instead of Redis, a vector DB, a memory framework, and a pile of MCP glue code.

Building an AI agent that actually remembers things usually means stitching together four separate systems: Redis for a fast KV cache, a vector database for semantic recall, a Mem0/LangMem-style memory framework on top of that, and a stack of custom MCP glue code so your agents can actually reach any of it. Four services, four bills, and a constant risk of the cache and the vector store quietly disagreeing with each other.

Cache-Pot is my bet that these aren't actually four separate problems — they're one memory engine with different retrieval modes. It's a single, Redis-compatible server where agents cache, remember, retrieve, share, and reason over information, and since it speaks the real Redis protocol, adopting it starts with swapping a connection string, not rewriting your stack.

	Redis	+ Vector DB	+ Mem0/LangMem	+ MCP adapters	Cache-Pot
Fast KV cache	✅				✅
Semantic/prompt/tool-call caching			partial		✅
Vector search		✅			✅
MCP-native tool access				✅	✅
Agent memory (semantic recall)			✅		✅
Shared memory across agents/models			partial		✅
Memory consolidation + knowledge graph			partial		✅
Multi-tenant isolation + memory versioning			partial		✅
Separate services to run & pay for	—	2	3	4	1

What's actually built (v0.7.0)

I'm being deliberately honest about status here rather than overselling it — the repo's ROADMAP.md tracks exactly what's real per version. As of v0.7.0, Cache-Pot is a genuinely working Redis-compatible cache (RESP2, core data structures, TTL, transactions, pub/sub), plus:

Semantic/prompt/tool caching — CACHE.SEMANTIC, CACHE.PROMPT, TOOL.CACHE
A native vector store — VECTOR.UPSERT / SEARCH / DELETE, no bolted-on Pinecone
Shared agent memory — MEMORY.*, AGENT.REMEMBER / RECALL
A native MCP server exposing that same memory to any connected agent
Observability and cost analytics — /metrics, /stats, /dashboard, eviction via --max-entries
Real LLM-backed consolidation and a knowledge graph — SUMMARY.CREATE, GRAPH.EXTRACT / GRAPH.RELATED
Enforced multi-tenancy and full memory version history — --workspace-credentials, MEMORY.HISTORY

Getting it running

docker compose -f deployments/compose/docker-compose.yml up --build
redis-cli -p 6380 PING
redis-cli -p 6380 SET hello world
redis-cli -p 6380 GET hello

Or build it directly:

go build -o bin/cachepotd ./cmd/cachepotd
./bin/cachepotd --port 6380

Because it's RESP2-compatible, it works with any existing Redis client — go-redis, redis-py, ioredis — for whatever subset of commands it currently implements.

Full docs (vision, command reference, architecture, and the honest state of "Redis-compatible") live in the repo's own VitePress site under docs/.