one store. seven floors. zero queries.
a real agidb store, a real WebSocket, a real tools/call JSON-RPC frame on every command. the right column explains what's happening at the substrate level. the left terminal prints the verbatim response payload — same wire shape claude desktop uses. install agidb →
-
floor 3 · observe 1/8
five episodes land as structured triples
"Sarah recommended Bawri" binds three concepts in the HDC index. text-only sentences still get a 8192-bit gist signature so they're recallable immediately — even before any extractor model is loaded.
read about floor 3 → -
floor 6 · goal 2/8
an active goal biases future recall
the goal "find a thai place" gets its own HDC signature. from now on every recall is goal-conditioned — matches aligned with the active goal float up. substrate primitive, not text inside an episode.
read about floor 6 → -
floor 6 · belief 3/8
first-class revisable beliefs
"Sarah likes thai food" with confidence 0.80 and two citing episodes. beliefs are revised, never overwritten — every change writes a row to an append-only revision log you can replay.
read about floor 6 → -
layer 1 · tier-A recall 4/8
the cascade: exact → similarity → gist → nearest
three "Sarah recommended Bawri" episodes hit tier-A exact in 0ms. no embedding API call, no vector DB, no LLM in the read path — the entire retrieval is hypervector hamming + an in-memory concept index.
read about the cascade → -
phase 6 · consolidation 5/8
sleep: cluster, bundle, contradict
three identical "Sarah recommended Bawri" episodes get bundled into a single semantic atom. the original episodes stay intact — the atom is a derived index, not a destructive merge.
read about consolidation → -
floor 6 · revise_belief 6/8
contradicting evidence drops confidence
a new episode contradicts "Sarah likes thai". the belief drops 0.80 → 0.70, the revision log captures it, and
read about belief revision →what_do_i_believe("Sarah")returns the live value. nothing is overwritten. -
floor 7 · self-vector 7/8
the agent gets a slowly-drifting identity
an 8192-bit self-vector EMA-updates toward the bundle of consolidated atoms after every pass. "what kind of agent am I right now" becomes queryable, inspectable, and subtractable.
read about floor 7 → -
article XVI · unlearn 8/8
non-destructive cascading forget
unlearn(Concept(Sarah)) cascades: 6 episodes tombstoned, 1 belief auto-corrected, the self-vector subtracts the forgotten content (hamming weight drops 2060 → 119), and a permanent
read about the unlearn primitive →LearningEvent::Unlearnedaudit row is written. recoverable for 30 days, then compaction-eligible. audit never expires.