Knowledge systems

Human judgment as a learning system

Route novel decisions to experts; compile resolved decisions into reusable standards.

From candidate to reusable standard

The inbox and outbox are one stage in a loop that learns from expert judgment.

Implementation

Try Engram with Codex
Agent fit

First-class Codex support over MCP; no agent switch required.

Install

brew install gentleman-programming/tap/engram

Connect

engram setup codex

What it adds

MCP tools, the Memory Protocol, and compaction recovery instructions.

Official setup guide ↗
Text model
Assisted

Engram · Capture

A candidate arrives with its evidence, scope, uncertainty, and the decision it needs. This is the human queue's inbox.

Artifact observer Agent-mediated Lifecycle adapter Coupled runtime
Engram implementation

Agents explicitly save memories; hooks observe session boundaries, not raw coding activity.

What happens

Raw events acquire timestamps, project and user scope, source provenance, and links to the task that produced them.

Design choices

Capture everything or only explicit saves; synchronous hooks or background observation; transcripts, diffs, outcomes, and user corrections.

Failure if wrong

Important experience disappears—or noisy transcripts become permanent truth before their value is known.

Diagram model
Coding sessionOBSERVABLE SIGNALS EngramSTAGE TRANSFORMATION Durable handoffNEXT STAGE INPUT Prompt Tool calls Code changes Outcomes Scoped event stream Candidate experience capture attach provenance
Memory formsepisodicworkingActionscreate
Catalog lens: compare automatic capture, coding-tool hooks, and provenance.
Rank all 18 memory systemsLeaders: Engram 8.77 · Remnic 8.71 · AgentMemory 8.54 · ProjectMem 8.49