AI infrastructure & developer tooling
Agent Memory Hub.
Context that survives a change of machine, tool, or session.
I built a shared memory and context platform for an AI development fleet: canonical history, bounded native delivery, live recall, and verified routes to the work.
Where the work happens
- Agent memory files
- Project instructions
- Source provenance
Windows · macOS · Linux
Memory with history
- Versioned evidence
- Reviewable changes
- Scoped access
Collect · compile · recall
Native context
Bounded indexes in supported agent files.
MCP + CLI
Full detail, history, and verified repository routes.
Before checkout workRepository identity → placement policy → live verification
- My role
- Creator & engineer
- Focus
- Agent memory · context · routing
- Status
- Private tool · in use
The problem
Stop rebuilding the same context.
A decision made on one computer can be missing on the next. A coding agent’s context is often trapped inside a tool’s local files, while the repository, operating rules, and source history live elsewhere.
Agent Memory Hub collects eligible memory into one versioned store. A bounded index goes into supported native agent files; full detail stays available through MCP and CLI. The agent gets a useful starting point without loading the entire history into every session.
The work
A small starting context. The full history when needed.
Canonical memory with provenance
Keep facts and their source history together. Versions, lifecycle states, and reason-coded changes make it possible to review how context arrived and how it changed.
Native files plus live recall
Supported Claude Code and Codex targets receive bounded native context. MCP and CLI provide detailed recall and operational capabilities when the task calls for them.
Route before acting
Resolve a repository by identity and purpose, apply the operator’s placement policy, and verify the exact checkout. A familiar directory name is not enough to select where work belongs.
How it fits together
From a local fact to useful fleet context.
- 01
Collect & version
Gather eligible memory from registered machines and preserve source provenance and prior versions in the canonical store.
- 02
Compile & deliver
Build a context slice that fits the target. Preview writes, preserve surrounding content, back up existing files, and verify the result.
- 03
Recall & apply
Use the compact context to orient the agent, then retrieve detail and verify repository placement before taking action.
Native context delivery and runtime recall are complementary paths. A compact file is a starting index, not a claim that every fact fits in a prompt.
Engineering decisions
Make context reviewable and access explicit.
One canonical writer
The hub owns canonical memory. Hub-driven collection and delivery reduce the ambiguity of competing copies and keep reach separate from an agent’s local session.
Bounded native delivery
Context has a capacity limit. Compact indexes and source-aware recall preserve access to history without pretending that a large file will always be loaded in full.
Guarded changes
Previews, backups, writeback checks, retained history, and explicit no-write outcomes make a failed or ineligible delivery visible rather than silently successful.
Scoped service capabilities
Approved capabilities let an agent use selected services without receiving vault values. Memory can refer to a credential without becoming the credential store.
- Go
- MCP + CLI
- Native agent context
- SSH fleet collection
- Content-addressed history
- Policy-backed routing
What this demonstrates
Memory becomes useful when it connects to the work.
AMH brings context, source history, repository placement, and scoped capabilities into a common operational layer, while preserving the distinctions between what an agent knows and what it may do.