Skip to main content
Glama

mcp-memory

Smart memory for AI agents. Memories decay, topics are frequency-weighted, one-time questions don't become obsessions.

Solves the Karpathy problem: "A single question from 2 months ago keeps coming up as a deep interest with undue mentions in perpetuity."

What's New in v0.2.0

  • Auto-categorization — no need to specify category, inferred from content

  • Semantic dedup — bigram similarity prevents duplicate memories

  • Preference supersede — "prefers dark mode" then "prefers light mode" updates, not duplicates

  • Recall auto-reinforces — searching for a topic counts as a mention

  • Recall auto-prunes — dead memories cleaned up on every read

  • System prompt injection — active memories provided via MCP prompts capability

  • Fuzzy forget — "VS Code" matches "User prefers VS Code for all editing"

  • 7 tools → 4 tools — simpler API, higher adoption (backwards compatible)

Related MCP server: engram-mcp

In Action

Day 1: User asks 5 questions (Rust, dark mode, Python, job title, Haskell)

  #1 [ACTIVE] rel=1.000 cat=preference "User prefers dark mode in all editors"
  #2 [ACTIVE] rel=0.900 cat=fact       "User works as a senior software engineer"
  #3 [FADING] rel=0.500 cat=question   "User is building a Python web scraper"
  #4 [FADING] rel=0.300 cat=one-time   "User asked about Rust programming"
  #5 [FADING] rel=0.300 cat=one-time   "User asked what Haskell monads are"

Day 2-5: User mentions Python 4 more times → auto-upgraded to "interest"

  #1 [ACTIVE] rel=2.658 mentions=5 cat=interest    "Python web scraper"
  #2 [ACTIVE] rel=1.000 mentions=1 cat=preference  "dark mode"
  #3 [ACTIVE] rel=0.900 mentions=1 cat=fact         "senior software engineer"
  #4 [FADING] rel=0.300 mentions=1 cat=one-time     "Rust" ← FADING, won't obsess
  #5 [FADING] rel=0.300 mentions=1 cat=one-time     "Haskell" ← FADING, won't obsess

After 60 days:
  Rust:   0.3 × 0.5^(60/7) = 0.0008 → DEAD (gone, as it should be)
  Python: 0.8 × 0.5^(60/60) × 3.32 = 1.329 → STILL ACTIVE (real interest)

How It Fixes This

Current LLM Memory

mcp-memory

Ask about Rust once → mentioned forever

Ask once → fades in 7 days

All memories equal weight

Categories: one-time (7d), question (14d), interest (60d), preference (180d)

No decay

Exponential decay — old memories naturally fade

No frequency tracking

Mentioned 5+ times → auto-upgrades from "question" to "interest"

Keyword matching

Bigram similarity + relevance scoring

Agent must decide to remember

Auto-categorizes from content patterns

Contradicting preferences coexist

New preference supersedes old one

Manual cleanup required

Auto-prunes dead memories on recall

Install

"mcpServers": {
  "memory": {
    "command": "npx",
    "args": ["-y", "mcp-memory"]
  }
}

Tools

Tool

What it does

remember

Store a memory. Auto-categorizes from content. Auto-deduplicates via bigram similarity. Supersedes conflicting preferences.

recall

Retrieve memories ranked by relevance. Auto-reinforces top match. Auto-prunes dead memories.

forget

Delete a memory by ID or fuzzy content match.

inspect

Debug view: all memories with decay status, relevance scores, category breakdown, health.

Auto-Categorization

No need to specify category — it's inferred from content:

Content Pattern

Auto-Category

Decay

"prefers X", "likes X", "always uses X"

preference

180 days

"works as X", "is a X", "lives in X"

fact

365 days

"actually X", "meant X", "wrong"

correction

365 days

"currently building", "working on"

context

30 days

"what is X", "how to X"

one-time

7 days

anything else

question

14 days

You can still override: remember(content: "...", category: "preference")

Examples

Auto-categorized preference:

remember(content: "User prefers TypeScript over JavaScript")
→ Auto-detected as "preference". Persists 180 days.

Semantic dedup:

remember(content: "Works as data scientist at Google")
remember(content: "Works as senior data scientist at Google")
→ Second call reinforces first (80% similar). Keeps longer version.

Preference supersede:

remember(content: "User prefers dark mode")
remember(content: "User prefers light mode")
→ Superseded: "dark mode" → "light mode". One memory, not two.

Recall auto-reinforces:

recall(query: "MCP servers")
→ Returns matching memories AND counts this as a mention.
  mention_count goes from 1 → 2 automatically.

Fuzzy forget:

forget(content: "VS Code")
→ Matches and removes "User prefers VS Code for all editing"

The Math

relevance = base_weight × decay × frequency_boost

where:
  base_weight  = category-specific (0.3 for one-time, 1.0 for preference)
  decay        = 0.5 ^ (age_days / halflife_days)
  freq_boost   = 1 + log2(mention_count)

A one-time question from 2 months ago: 0.3 × 0.5^(60/7) × 1.0 = 0.0003 → effectively zero. Won't surface.

A preference mentioned 8 times, last week: 1.0 × 0.5^(7/180) × 4.0 = 3.89 → top of every recall.

Backwards Compatibility

v0.2.0 still accepts the old v0.1.0 tool names (reinforce, prune, stats). They map to the new tools internally. No breaking changes.

License

MIT

Available Tools

4 tools
forgetA

Delete a memory by ID or content match (fuzzy). Use when the user says to forget something.

Inputs: id, content.

inspectA

Debug view: all memories with decay status, relevance scores, ages, and health stats. Also shows category breakdown.

Inputs: no named parameters.

recallA

Search memories by topic. Auto-reinforces strong matches. Auto-prunes dead memories. Leave query empty to get all active memories ranked by relevance.

Inputs: query, limit, min_relevance.

rememberA

IMPORTANT: Call this whenever the user reveals a preference, fact about themselves, correction, or recurring interest. Auto-categorizes if no category given. Auto-deduplicates similar content. Categories: one-time (7d), question (14d), interest (60d), preference (180d), correction (365d), fact (365d), context (30d).

Inputs: content (required), category, tags.

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4.2/5.0

    Scored across 4 tools

    Disambiguation5/5

    Each tool targets a distinct operation on memories: forget deletes, inspect shows debug view, recall searches, and remember creates. There is no overlap or ambiguity.

    Naming Consistency5/5

    All tool names are single-word lowercase verbs (forget, inspect, recall, remember), following a consistent imperative pattern.

    Tool Count5/5

    With 4 tools, the server covers the essential CRUD-like operations for a memory system (create, read, delete, and inspect) without being excessive or insufficient.

    Completeness4/5

    The tool set covers creation, search, deletion, and inspection, but lacks an explicit update tool. While auto-reinforcement and categorization handle some updates, a direct edit capability is a minor gap.

    Maintenance

    ActivityInactive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      Not graded
      quality
      D
      maintenance
      Persistent memory for AI agents — organized by time and space. Important memories get promoted, noise decays naturally, and related knowledge clusters into a browsable topic tree. Fully automatic.
      28
      MIT
    • A
      license
      A
      quality
      C
      maintenance
      Persistent semantic memory for AI agents. SQLite-backed, local-first, zero config. Semantic search via Ollama embeddings with keyword fallback. Tools: remember, recall, history, forget, stats.
      17
      37 npm
      1
      MIT
    • A
      license
      C
      quality
      F
      maintenance
      Cognitive memory for AI agents. Implements Atkinson-Shiffrin multi-store memory (sensory → STM → LTM), semantic RAG with fastembed, Ebbinghaus forgetting curves, trust scoring, and metacognitive guard. 76+ MCP tools. 100% local, MIT licensed.
      100
      27
      AGPL 3.0
    • A
      license
      Not graded
      quality
      B
      maintenance
      Open-source AI memory layer for LLM agents. Importance scoring, temporal decay, hierarchical memory (facts, summaries, themes), YMYL prioritization, and active retrieval with contradiction detection. Supports OpenAI, Anthropic, Ollama. Local-first with SQLite + FAISS.
      50
      Apache 2.0