Your AI keeps forgetting you — brew its memory instead
Your AI keeps forgetting you — brew its memory instead
introduce yourself once. set the rules once. your AI remembers — next week, next project, next agent
the repo, front and center → github.com/TencentCloud/TencentDB-Agent-Memory
setup in one breath: clone → one .env file → ./start-all.sh → paste the one-liner into your AI · full guide: docs
jar 1 · 
chat memory
your preferences, facts and decisions survive between sessions — “I always want short answers” is said once
each agent only sees the memories you allow it to see
jar 2 · 
skills
when your AI finishes something tricky, the working recipe gets bottled — with versions and trigger rules, not just a prompt paste
any other agent can pick up that skill on day one instead of rediscovering it
jar 3 · 
LLM-wiki
your documents turn into a wiki every agent can consult — no more re-reading the PDF from page one every session
jar 4 · 
code-graph
your codebase becomes a map — “don’t refactor the auth module, mobile still uses it” is remembered forever, not re-explained weekly
how the brewing works
raw chats are distilled layer by layer: L0 conversation → L1 atoms (facts) → L2 scenarios (recipes) → L3 persona (who you are)
the result is a human-readable save file — a brand-new agent loads your whole history in one call
hookup is one URL swap: Claude Code, Codex, DeepSeek Harness, OpenClaw, OpenCode — no plugins needed · your control panel: localhost:8125
the memory garden — pick your flavor, all free
- Mem0 (68.9K ) — the simplest add-and-remember layer · tested at 92.5% recall with ~7K tokens instead of 25K
Graphiti (32.8K ) — time-travel memory: facts keep “was true then / true now” windows · game NPCs finally hold grudges properly
Cognee (33.2K ) — docs + code + chats woven into one knowledge graph
Basic Memory (4.6K ) — memory as plain Markdown you can read, edit and back up yourself — the no-black-box pick
Letta / MemGPT (25.4K ) — the agent that manages its own brain, rewriting its memory and skills
PMB — a tiny SQLite memory for coding agents, no LLM needed to read it
▸ one minimized window (click to open)
the receipts — long-session tests report −61% tokens on WideSearch, SWE-bench success 58→64%, PersonaMem 48→76% (vendor-published; an issue caught a math typo and it was fixed in public — that’s how open source should behave)
the “vector-DB killer” defused — it kills the mandatory separate vector-DB service (SQLite + keyword search by default, optional embeddings) — it doesn’t ban vectors, it de-gatekeeps them
privacy, honestly — “local-first” means the memory stays home, but extraction still talks to whatever model you point it at — use Ollama for a fully-local loop
fun use — agents sharing a family brain: appliance manuals, allergies, school calendars — with private vs family visibility controls
▸ the fine print, two lines
setup = Docker + two model configs — one brain for extracting memories, one for serving your agent (a cheap one for memory, a strong one for work)
never pour secrets — banking, passwords and private messages don’t belong in any AI’s memory jar
Your AI always had the brains. Now it finally keeps them.
!