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.
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.
Source: https://github.com/digitalforgeca/sulcus (public SDK + integrations)