A multi-agent swarm memory hub. The substrate that oversees, manages, and learns from a swarm's shared memory — built on one rule: the model is the battery, so confine it to write time and make read time fully mechanical. Model writes, grep reads, git remembers.
⚡ The battery, named. In any self-evolving memory system the LLM is the battery — you can't remove it, because deciding what to keep and abstracting patterns into skills need judgment. TOPH CORTEX moves the battery to write time only. The read path never calls a model, so the corpus is usable with the power out. Anyone claiming a memory system needs no model is hiding the battery.
the loop
◈ RECALL
grep + BM25 over the memory. Deterministic, offline.
0 model calls
→
▸ ACT
the agent does the task, already knowing what the swarm learned.
→
✎ REMEMBER
episodic appends; semantic stages a diff (never straight to main).
battery: write time
→
⛨ GATE
validation / review before commit — the Closure Loop.
→
⎇ git
git remembers. git log = a memory's maturity curve, free.
try the read path — zero LLM, right here
BM25 over the embedded seed memory — the same deterministic algorithm as bin/recall.py, no model, in your browser. Same query → same results forever.
the substrate — plain files in git
semantic/facts, one .md/topic — shared truth
episodic/append-only logs, per agent — cite [[topic]]
Every node human-readable, diffable, useful with the power out.
the referee — deterministic, not a nested agent
hit-rate = cited by a later session; ★ = cited by ≥2 distinct agents (earned its keep swarm-wide). Staleness is flagged, never silently rewritten. Live snapshot from bin/referee.py.
David Lee Wise's real git agents, pointed into the hub. Each links to its git repo and its sphere on the ud0 corpus (live at 0root.ai). Registered through the real write path — bin/recall.py "provenance monk" finds nom cold.
oversees · manages · learns
OVERSEES — the whole swarm's memory, staged changes, staleness, maturity. MANAGES — recommends which mature facts to promote past the gate, flags the stale & the cold, surfaces consolidation candidates. Promotion itself is a human / bin/promote.py action — never the overseer's, so the gate holds. LEARNS — and here is the honest part:cortex.py --learn appends a snapshot to cortex_state.json recording which memories/skills have earned their keep. No weights, no battery in the learning — it is a deterministic, git-backed record the whole swarm can trust cold.
run it
# READ — mechanical, no model, offlinepython bin/recall.py "battery read path"
# WRITE — episodic appends; semantic STAGES a diff (the gate)python bin/remember.py --agent me --session s5 --episodic --item "cited [[substrate]]"
# CONSOLIDATE — mechanical cluster, then STAGE a skill skeleton per cluster for the battery to complete (a separate deliberate call, not this script)python bin/consolidate.py# GATE — validate a staged fact + promote it into main memory (deterministic, no model)python bin/promote.py onboarding
# REFEREE + OVERSEERpython bin/referee.py · python bin/cortex.py --learn
# and a SessionEnd/Stop hook fires the mechanical half automatically.
The honest limit, stated plainly. You still need a model for the two genuinely cognitive operations — deciding what's worth keeping, and abstracting patterns into skills. No architecture eliminates that; anyone claiming otherwise is hiding the battery. What this guarantees instead: the artifact never depends on the battery after commit — internet dies, you still have a greppable, verifiable, git-lineaged corpus the next session, the next model, or you with your eyeballs can use cold. LIT the tools, the substrate, and every metric are a real running system (stdlib, offline, end-to-end tested) · FIG this dashboard is a snapshot of the seed / demonstration swarm — the alpha/beta/gamma logs are authored fixtures showing the shape; only agents that actually run the tools generate live entries (delta's onboarding was, and it was validated through the gate). Which is, not coincidentally, just AKASHA with hooks bolted on.