AgentPM™

memory#

Build and inspect kind: "memory" packages.

Overview#

agentpm memory is the CLI surface for Memory Blueprints. It currently supports:

  • agentpm memory build to validate a blueprint and generate resolved contracts
  • agentpm memory inspect to inspect a local or installed Memory package without rewriting it

The command group is diagnostic and packaging-oriented. It does not create a live store, bind a blueprint to a backend, persist records, execute lifecycle operations, or evaluate triggers.

Commands#

Build#

agentpm memory build [PATH] [--manifest <path>]

Use build when you want AgentPM to:

  • validate the authored blueprint and referenced source schemas
  • generate one resolved contract for every declared space-and-record-type pairing
  • write memory/contracts/index.json
  • write memory/build.json

Typical flow:

agentpm lint --strict
agentpm memory build
agentpm publish --dry-run

Inspect#

agentpm memory inspect <PATH_OR_PACKAGE> [--json]

Target resolution order:

  1. existing local directory
  2. explicit local agent.json
  3. installed Memory package reference such as @namespace/name@range

Accepted installed-package forms:

  • @namespace/name
  • @namespace/name@0.1.0
  • @namespace/name@^0.1
  • memory:@namespace/name@0.1.0

inspect is read-only. It never rebuilds or repairs package state.

Build outputs#

A successful build writes:

memory/
  build.json
  contracts/
    index.json
    <space>.<record_type>.schema.json

memory/contracts/index.json#

The contract index inventories the generated contracts and includes, per entry:

  • space
  • record_type
  • schema_version
  • model
  • source_schema
  • generated path
  • generated contract hash

memory/build.json#

The build metadata records portable authored-input and generated-output hashes so AgentPM can detect:

  • never-built state
  • stale authored input
  • missing generated output
  • modified generated output
  • unsupported metadata
  • inconsistent metadata

Example: build a blueprint#

agentpm init --kind memory --name conversation-continuity --description "Durable memory contract for a support assistant"
cd conversation-continuity
agentpm lint --strict
agentpm memory build

Sample result:

Memory build: conversation-continuity@0.1.0
Scopes: 1
Record types: 1
Spaces: 1
Operations: 0
Contracts written: 1
Output: memory/contracts

Example: inspect a local package#

agentpm memory inspect .

Typical text output includes:

  • package identity and paths
  • normalized build status such as fresh, not_built, stale, invalid, or unsupported
  • authored scopes, record types, spaces, and lifecycle operations
  • generated contract inventory
  • structured mismatch details when generated output is stale or malformed

Machine-readable output:

agentpm memory inspect . --json

Publish readiness#

agentpm publish does not build Memory output for you. A Memory package must already have fresh generated files.

If authored inputs changed since the last build:

agentpm memory build

Then try publish again:

agentpm publish --dry-run

Common failure cases#

Not built yet#

Memory Blueprint is not built.
Run:
agentpm memory build

Authored inputs changed#

inspect reports a stale state and identifies the changed source schema, manifest, or generated file where possible.

Modified or missing generated output#

inspect reports an invalid state and lists the affected generated files without rewriting them.

SDK boundary#

Memory Blueprints become useful to SDK consumers after install:

  • Node: loadMemory() and loadMemoryContract()
  • Python: load_memory() and load_memory_contract()

Those SDK calls are metadata and contract loaders only. They do not provide live record CRUD or a hosted memory runtime.

←CLIknowledge
→CLIwhoami