@leonard-memento/core

Amnesia insurance for coding agents: a curated, Markdown project memory index injected into context compaction, so auto-summarization never forgets that memory exists. Zero dependencies.

Part of the leonard-memento family — this package carries the shared logic and the memento CLI; the per-agent plugins build on it.

CLI

memento path [--project DIR]    # absolute memory index path (exit 1 if absent)
memento inject [--project DIR]  # the injection block (silent, exit 0 when absent)
memento init <agent>            # idempotent per-agent setup
memento init --list             # supported agents

API

• projectHash(directory) — first 16 hex chars of sha256 of the canonical project path (~-relative input throws: expand it yourself).
• memoryIndexPath(directory, options?) — absolute index path.
• readProjectIndex(directory, options?) — the frozen injection block, or null when the project has no memory yet.
• resolveMemoryRoot(overrides?) — effective memory root; never throws.
• initAgent(projectDir, agent) / TARGETS — the init machinery.

Options and their env fallbacks: memoryRoot / MEMENTO_MEMORY_ROOT, indexFile / MEMENTO_INDEX_FILE.

Memory convention

One flat Markdown tree, outside the repo:

~/.config/memento/projects/<hash16>/MEMORY.md   # the index
                                  ├── topic-a.md
                                  └── topic-b.md

MEMORY.md is one line per fact: - [[name]] — short description. Fact files carry front-matter (name, description, type, dates). The legacy root ~/.config/opencode/memento/projects keeps resolving, so existing opencode-memento v1.0.0 trees work unchanged.