edit: Apologies, I can't believe I put in the wrong url suffix. We're @ https//sulcus.dforge.ca
Hi HN,
AI agents forget everything between sessions. The standard fix is a vector DB — store embeddings, retrieve on similarity. But that's passive. The agent has no control over what stays, what fades, or what matters.
We built Sulcus to fix this. It's a memory system where the agent controls its own memory lifecycle — and the memory system can react autonomously.
The novel bit: reactive triggers.
SULCUS | The Virtual Memory Management Unit for AI Agents
SULCUS provides thermodynamic context management for LLMs, reducing token burn by 90% via intelligent memory paging.
You define rules on the memory graph:
on_store + pin → auto-pin every preference memory (immune to decay)
on_recall + boost → reinforce memories every time they're searched (automated spaced repetition)
on_decay + notify → alert the agent before critical knowledge goes cold
on_threshold + webhook → fire HTTP callbacks when memory heat crosses a boundary
Triggers fire automatically during normal operation. When one fires, the agent sees trigger_notifications inline in the tool response. No polling, no cron — the memory system talks back.
No other memory system does this. Mem0, Zep, Letta — all passive stores. Sulcus triggers are reactive.
The engine underneath:
Memories have thermodynamic properties — heat (relevance), decay profiles per type (episodic fades fast, procedural persists), stability from repeated recall, configurable half-lives. A background tick applies decay and fires threshold/decay triggers. 30+ tunable knobs exposed via API.
Stack:
Rust core (sulcus-local) with embedded PostgreSQL — runs as MCP sidecar
29 MCP tools (Claude Desktop, OpenClaw, any MCP host)
Cloud sync via CRDT replication across agents/machines
Zero-dependency SDKs: npm install sulcus / pip install sulcus (coming)
REST API + dashboard at sulcus.dforge.ca
We want feedback on the trigger model. Are there events/actions we're missing? Does the thermodynamic metaphor resonate or confuse? We're also curious if anyone has tried reactive memory in production agents.
mcdoolz•1h ago
Hi HN,
AI agents forget everything between sessions. The standard fix is a vector DB — store embeddings, retrieve on similarity. But that's passive. The agent has no control over what stays, what fades, or what matters.
We built Sulcus to fix this. It's a memory system where the agent controls its own memory lifecycle — and the memory system can react autonomously.
The novel bit: reactive triggers. SULCUS | The Virtual Memory Management Unit for AI Agents SULCUS provides thermodynamic context management for LLMs, reducing token burn by 90% via intelligent memory paging. You define rules on the memory graph:
on_store + pin → auto-pin every preference memory (immune to decay) on_recall + boost → reinforce memories every time they're searched (automated spaced repetition) on_decay + notify → alert the agent before critical knowledge goes cold on_threshold + webhook → fire HTTP callbacks when memory heat crosses a boundary
Triggers fire automatically during normal operation. When one fires, the agent sees trigger_notifications inline in the tool response. No polling, no cron — the memory system talks back.
No other memory system does this. Mem0, Zep, Letta — all passive stores. Sulcus triggers are reactive.
The engine underneath:
Memories have thermodynamic properties — heat (relevance), decay profiles per type (episodic fades fast, procedural persists), stability from repeated recall, configurable half-lives. A background tick applies decay and fires threshold/decay triggers. 30+ tunable knobs exposed via API.
Stack: Rust core (sulcus-local) with embedded PostgreSQL — runs as MCP sidecar 29 MCP tools (Claude Desktop, OpenClaw, any MCP host) Cloud sync via CRDT replication across agents/machines Zero-dependency SDKs: npm install sulcus / pip install sulcus (coming) REST API + dashboard at sulcus.dforge.ca
We want feedback on the trigger model. Are there events/actions we're missing? Does the thermodynamic metaphor resonate or confuse? We're also curious if anyone has tried reactive memory in production agents.
Source: https://github.com/digitalforgeca/sulcus (public SDK + integrations)