Memora

Memora — Persistent memory and knowledge management

Memora is an MCP memory server. Use it at session start to load context, when the user asks about past work or stored knowledge, and when saving important information for future sessions.

When to invoke

• Session start: load relevant memories for the current task
• User asks about past work, decisions, or stored knowledge
• Saving important findings, decisions, architecture notes, or research
• User explicitly asks to remember, recall, or search memories

Tool usage guidelines

Retrieval — lineage defaults are enforced

Defaults (do not re-state them on every call):

• memory_list / memory_semantic_search / memory_hybrid_search default to follow="active" — superseded memories are excluded.
• memory_get defaults to follow="latest" — a superseded id resolves to the current leaf of its chain.

memory_list(tags_any=["memora/todos"])
memory_semantic_search(query="cloud backend")
memory_get(memory_id=157)  # returns the current version if 157 was superseded

When you need non-default lineage:

# Resolve hits to current leaves (dedupe versions in a search)
memory_semantic_search(query="roadmap", follow="latest")

# Full supersession chain
memory_get(memory_id=170, follow="full_history")

# Forensic / unfiltered — EXPLICIT only. Omitting follow is NOT unfiltered.
memory_list(follow="all")
memory_semantic_search(query="...", follow="all")
memory_get(memory_id=157, follow="all")  # exact id, no chain walk

Do not cargo-cult follow="active" on every call — that is already the list/search default. Pass follow only to change behaviour (latest, full_history, or all).

Saving knowledge — use memory_absorb

Prefer memory_absorb over memory_create when saving knowledge. Absorb automatically checks for duplicates, supersedes outdated memories, links related ones, and consolidates related new facts into single richer memories.

Write detailed, context-rich facts — not tiny one-liners. Each fact should be a full sentence or short paragraph with enough context to be useful on its own. Related facts passed together are automatically merged into a single consolidated memory via LLM synthesis.

memory_absorb(
    facts=[
        "clmux v0.4.16.1 fixes hidden pane text leakage by restricting tmux allow-passthrough to only the visible TUI window instead of globally, preventing hidden reviewer and parking windows from leaking escape sequences",
        "clmux sidebar now filters out _reviewers and parking windows from list-panes queries, and the ! notification indicator persists across workspace switches until the user responds"
    ],
    source="manual",
    tags=["clmux", "bugfix"]
)

Absorb handles dedup and consolidation automatically:

• Duplicate → skipped (no new memory created)
• Update → creates new memory + supersedes the old one
• Contradiction → creates new memory + links with contradicts edge
• Related → creates new memory + links with related_to edge
• New → creates new memory (no matches found)
• Consolidated → related new facts merged into a single richer memory

Use dry_run=True to preview what absorb would do without writing.

Use memory_create directly only for:

• Raw/unprocessed content you want stored verbatim
• Structured entries (TODOs, issues, sections) via memory_create_todo, memory_create_issue, memory_create_section

Updating memories

Use memory_update only for corrections (typos, metadata fixes) — not for evolving knowledge. For evolving knowledge, use memory_absorb — it handles supersession automatically.

Linking

Use typed edges to express relationships:

• supersedes — new version replaces old (enables lineage walking)
• contradicts — conflicting information (flag for resolution)
• implements — concrete implementation of a plan/design
• extends — builds upon existing knowledge
• references — general reference/citation
• related_to — loose association

Search strategy

1. Start with memory_semantic_search for conceptual queries (follow="active")
2. Use memory_hybrid_search when you need both keyword and semantic matching
3. Use memory_list with tag/metadata filters for structured browsing
4. Use fields parameter to reduce response size: fields=["id", "content_preview", "tags"]
5. Use content_mode="preview" (default) for scanning, content_mode="full" only when you need complete content

Context efficiency

• Default limit=20 on list. Use limit=-1 only when you truly need everything.
• Use fields projection to fetch only what you need.
• Prefer content_mode="preview" (default) over full content for scanning.
• Use memory_get with specific IDs after finding relevant results via search.