Amnesia
Yesterday's decisions, your stack preferences, the unwritten rules of the project — gone when the session ends.
So it recommends the approach you already rejected, and asks what you answered three times.
AllMyMind · 记忆森林
A leaf is something you said. A branch is how it connects. A tree is a topic that has grown. What should fall, falls — because it forgets clearly.
Close the context window and everything is gone. Paste the whole history in and attention dilutes. A window is not a memory.
Yesterday's decisions, your stack preferences, the unwritten rules of the project — gone when the session ends.
So it recommends the approach you already rejected, and asks what you answered three times.
Cramming history into the prompt inflates tokens and drowns the one line that mattered.
Costs go up, quality goes down.
Even when it remembers, it cannot say why this came up. Nothing to debug, nothing to reuse.
Unexplainable memory cannot be trusted — or improved.
AllMyMind is not a wrapper around a vector store. It answers “what should memory look like” before “where do we put it”.
Memory is not a pile of loose vectors. Trunks are themes, branches are threads, leaves are what you actually said.
Grow the tree first, then look for leaves.
Grass untouched for two days withers; shrubs for four. Leaves fall into the fallen layer — still findable, but out of the way.
You have to know what to forget before you know what is worth keeping.
Every “I remember” can name the branch it hangs on, when it was stored, and why it surfaced now.
Memory matters when it can be questioned.
Leaves, branches, trees, shrubs, grass — each has explicit rules. Code decides what shape it becomes.
One sentence, one decision, one excerpt. Lands in the pending layer, waiting to be placed.
Three solid branches and three leaves each make a tree. If it cannot hold, it falls back to leaves.
It searches meaning, not keywords: “remote access fails after locking” finds the RustDesk headless note.
Drop a whole folder in. Inspect first: structural reading costs zero model calls, garbage is refused.
Attached to trunks only, excluded from search. Not even “there is a memory here” is revealed.
Elsewhere, it offers a card first: how many, how long, last stored. You approve before any text is sent.
Data lives in local SQLite. Models are used only for digging and grouping — it runs without a cloud.
MCP needs PYTHONPATH in the env block — see the docs.
Read the docsFour tables plus local Chroma; shape rules (tree / shrub / grass) enforced in code.
Manual pipeline: select a tree, “send to conversation”, packaged and copied to clipboard.
inspect → run → apply, with whole-batch rollback on failure.
AllMyMind is in beta. If you are building agents too, come try the first patch of forest.