MaxGraph Memory Protocol

Your agent forgets everything between sessions. This skill fixes that.

MaxGraph Memory Protocol is the operating discipline we run on our own agents every day. It turns a knowledge-graph memory server into reliable long-term memory: the agent always knows the project's decisions, open bugs and unfinished work — no matter how many sessions, context windows or compactions happened in between.

What the agent learns

• Session priming — one call at session start loads a budgeted context block: project facts, owner decisions (verbatim), open bugs, open tasks, past session outcomes. No more re-explaining the project every morning.
• Write immediately — facts are saved the moment they appear, never batched to the end of a conversation that might get compacted or crash.
• Atomic observations — one fact per record, dated when it matters. Searchable, supersedable, never a wall of stale text.
• Stale-fact hygiene — outdated facts (old prices, reversed decisions) are superseded, not duplicated: the old line leaves search results but stays in the archive.
• Open-loop tracking — explicit task and bug statuses, so «what's left?» is a query, not a guess.
• Heartbeat work-state — after each stage, one dated fact records «done X → next Y». Context loss costs nothing.
• Clean session close — one call writes the day's summary and links everything the session touched.
• Re-priming after compaction — when the harness compresses the conversation, the agent re-reads the graph and continues exactly where it stopped.

Why it works

Agents fail at memory in predictable ways: they batch writes to the end (and lose them), they duplicate entities, they can't tell stale facts from current ones. This protocol names each failure mode and gives the agent an exact, mechanical rule against it. It is model-agnostic and host-agnostic — any agent that can call the memory tools can run it.

Requirements

A running MaxGraph server (open source, self-hosted, one Docker container, MCP + REST + dashboard) — or any compatible knowledge-graph memory backend. The protocol itself is tool-name-agnostic.

Who it's for

Teams running agents across long-lived projects: the agent remembers architecture decisions, active bugs, deferred tasks and owner preferences — and picks up mid-sentence after a compaction, a crash, or a week away.