Durable memory in a world of ephemeral context — a file-based vault that working agents read and write, chat surfaces only read, and every session starts from.

AgentMemory is the harness's long-term memory: a plain-Markdown vault, synced through Google Drive, that outlives any single session or repo. It holds three kinds of thing — your standing preferences and conventions, per-project state, and lessons learned along the way — as one fact per file with YAML frontmatter, so a fresh session recalls what the last one knew without you re-explaining it.

The vault is a directory tree, not a database. Entries live under a few well-known regions, and access splits by who is asking:

Region	Holds	Loaded
standards/ (∪ the legacy memory/_always-load/ pen)	standing preferences, conventions, durable fixes	every session start
Projects/<name>/	per-project goals, plans, progress	on demand, by project
diagnostics/	health scorecards, dreaming reports, digests, dated audits	by the tooling, not the agent

Machine state — caches, cursors, orchestration state — is not a vault region at all: it lives outside the vault in the engine state directory (~/.local/state/agentm/, its own git repo the runner commits), so the vault stays the knowledge surface both audiences read.

Access is asymmetric by surface: the filesystem working agents you run directly (Claude Code, Antigravity) read and write the vault following your entry conventions; chat surfaces (Claude.ai, Claude Desktop) read and suggest entries but never write. A read-only vault lint and a personal-notes link-discovery audit keep the corpus on-spec — both surface fixes for you to apply, never editing an entry themselves.

• Device-Wide Substrate — the vault is the device-wide store. Memory spans every project because it lives outside any one repo, on the machine, not in the tree.
• Orchestration and Auto-Detection — the push/pull surfaces that read and write vault state at session start and phase boundaries.
• Phases — phase state (PLAN, progress) is written to the vault, not held in the conversation; that is what lets one phase per session stay coherent.

Detail:

• AgentMemory context payload — the block you paste into any agent surface to point it at the vault.
• Vault lint checks · Note relatedness signals — the audit catalogs.
• Use AgentMemory in any agent · Audit the vault — the how-to recipes.

Designs:

• MemoryVault — the write-primitives → recall → reflection build, part by part.
• Agent Memory Evolution V1→V4 · Memory-OS Architecture (V5) — where it came from and where it's going.

Architecture · Designs · Home