hmem — Humanlike Memory for AI Agents

AI agents forget everything when a session ends. hmem changes that.

Beta: hmem is functional and actively used in production, but APIs and file formats may still change. Feedback and bug reports welcome. Also the parameters I chose need to be tested and tweaked.

hmem is a Model Context Protocol (MCP) server that gives AI agents persistent, humanlike memory — modeled after how human memory actually works.

Born as a side project of a multi-agent AI system, hmem solves a real problem: when you work across multiple machines or sessions, your AI instances start from zero every time. They duplicate work, contradict previous decisions, and lose hard-won context.

hmem fixes this.

The Problem

When working across multiple PCs with AI coding agents, every new session was a fresh start. Agents had no knowledge of previous decisions, duplicated work, produced inconsistencies, and wasted tokens catching up.

Existing RAG solutions are flat — every memory fragment has the same abstraction level. The agent either gets too much detail and wastes tokens, or too little and loses nuance.

The Solution: 5-Level Humanlike Memory

hmem stores and retrieves memory in five nested levels of detail — mirroring how human memory works.

Level 1  ──  Coarse summary         (always loaded on spawn)
  Level 2  ──  More detail
    Level 3  ──  Deep context
      Level 4  ──  Fine-grained specifics
        Level 5  ──  Full verbatim detail

A freshly spawned agent receives only Level 1 — the broadest strokes. When it needs more detail on a specific topic, it makes a tool call to retrieve Level 2 for that entry. And so on, down to full detail.

Result: Agents load exactly as much context as they need — no more, no less.

How It Works

Saving Memory

After completing a task, an agent calls write_memory with tab-indented content. The indentation depth maps to memory levels — multiple entries at the same depth become siblings.

write_memory(prefix="L", content="Always restart MCP server after recompiling TypeScript
	Running process holds old dist — tool calls return stale results
	Fix: kill $(pgrep -f mcp-server)")

Loading Memory

On spawn, the agent receives all Level 1 summaries. Deeper levels are fetched on demand — by ID, one branch at a time.

read_memory()              # → all L1 summaries
read_memory(id="L0003")    # → L1 + direct L2 children for this entry
read_memory(id="L0003.2")  # → that L2 node + its L3 children

Each node gets a compound ID (L0003.2.1) so any branch is individually addressable.

Updating Memory

Entries can be updated without deleting and recreating them:

update_memory(id="L0003", content="Corrected L1 summary")
update_memory(id="L0003.2", content="Fixed sub-node text")
append_memory(id="L0003", content="New finding\n\tSub-detail")

update_memory replaces the text of a single node (children preserved). append_memory adds new child nodes to an existing entry.

Obsolete Entries

When an entry is outdated, mark it as obsolete — never delete it:

update_memory(id="E0023", content="...", obsolete=true)

Obsolete entries are hidden from bulk reads and replaced by a summary line at the bottom:

--- 3 obsolete entries hidden (E0023, D0007, L0012) — use read_memory(id=X) to view ---

They remain fully searchable and accessible by ID. Past errors still teach future agents what not to do — knowledge is never destroyed, only archived.

Memory Curation

A dedicated curator agent runs periodically to maintain memory health. It detects duplicates, merges fragmented entries, marks stale pointers, and prunes low-value content — a form of the Ebbinghaus Forgetting Curve.

Key Features

• Hierarchical retrieval — lazy loading of detail levels saves tokens
• True tree structure — multiple siblings at the same depth (not just one chain)
• Persistent across sessions — agents remember previous work even after restart
• Editable without deletion — update_memory and append_memory modify entries in place
• Obsolete flag — mark outdated entries as obsolete; hidden from bulk reads but still searchable — knowledge is never destroyed, only archived
• Favorite flag — mark any entry as [♥] to always see it with L2 detail, regardless of category
• Access-count promotion — the top-N most-accessed entries are automatically shown with L2 detail ([★])
• Effective-date sorting — entries with recent appends surface to the top (old P entries grow over time without losing their position)
• Token-efficient bulk reads — only the most recent L2 child is shown in bulk reads, with a "+N more" hint
• Per-agent memory — each agent has its own .hmem file (SQLite)
• Skill-file driven — agents are instructed via skill files, no hardcoded logic
• MCP-native — works with Claude Code, Gemini CLI, OpenCode, and any MCP-compatible tool

Quick Start

Option A: Install from npm (Recommended)

npx hmem-mcp init

That's it. The interactive installer will:

• Detect your installed AI coding tools (Claude Code, OpenCode, Cursor, Windsurf, Cline)
• Ask whether to install system-wide (memories in ~/.hmem/) or project-local (memories in current directory)
• Configure each tool's MCP settings automatically
• Create the memory directory and hmem.config.json

After the installer finishes, restart your AI tool and call read_memory() to verify.

Don't forget the skill files! The MCP server provides the tools (read_memory, write_memory, etc.), but the slash commands (/hmem-save, /hmem-read) require skill files to be copied to your tool's skills directory. See the Skill Files section below — it's a one-time copy-paste.

Coming from the MCP Registry? Run npx hmem-mcp init first — it configures your tools and creates the memory directory. Then copy the skill files as described below.

Option B: Install from source

git clone https://github.com/Bumblebiber/hmem.git
cd hmem
npm install && npm run build
node dist/cli.js init

Option C: Manual Setup (no installer)

If you prefer to configure everything yourself:

1. Install

npm install -g hmem-mcp

Or from source: git clone https://github.com/Bumblebiber/hmem.git && cd hmem && npm install && npm run build

2. Register the MCP server

Claude Code — global registration:

claude mcp add hmem -s user -- npx hmem-mcp serve \
  --env HMEM_PROJECT_DIR="$HOME/.hmem"

OpenCode — add to ~/.config/opencode/opencode.json (or project-level opencode.json):

{
  "mcp": {
    "hmem": {
      "type": "local",
      "command": ["npx", "hmem", "serve"],
      "environment": {
        "HMEM_PROJECT_DIR": "~/.hmem"
      },
      "enabled": true
    }
  }
}

Cursor / Windsurf / Cline — add to ~/.cursor/mcp.json (or equivalent):

{
  "mcpServers": {
    "hmem": {
      "command": "npx",
      "args": ["hmem", "serve"],
      "env": {
        "HMEM_PROJECT_DIR": "~/.hmem"
      }
    }
  }
}

Windows note: Use forward slashes or double backslashes in JSON paths.

3. Verify the connection

Fully restart your AI tool, then call read_memory(). You should see a memory listing (empty on first run is fine).

In Claude Code, run /mcp to check the server status.

Skill Files

Skill files teach your AI tool how to use hmem correctly. Copy them to your tool's global skills directory, then restart your AI tool.

After copying skills, fully restart your terminal and AI tool — skills are loaded at startup and won't appear in a running session.

Available skills

Copy skills to your tool

Find the skills directory in the installed package:

HMEM_DIR="$(npm root -g)/hmem-mcp"

If you cloned from source, the skills are in the skills/ directory.

Claude Code:

for skill in hmem-read hmem-write hmem-save hmem-config hmem-curate; do
  mkdir -p ~/.claude/skills/$skill
  cp "$HMEM_DIR/skills/$skill/SKILL.md" ~/.claude/skills/$skill/SKILL.md
done

Gemini CLI:

for skill in hmem-read hmem-write hmem-save hmem-config hmem-curate; do
  mkdir -p ~/.gemini/skills/$skill
  cp "$HMEM_DIR/skills/$skill/SKILL.md" ~/.gemini/skills/$skill/SKILL.md
done

OpenCode:

for skill in hmem-read hmem-write hmem-save hmem-config hmem-curate; do
  mkdir -p ~/.config/opencode/skills/$skill
  cp "$HMEM_DIR/skills/$skill/SKILL.md" ~/.config/opencode/skills/$skill/SKILL.md
done

MCP Tools

Memory Tools

Curator Tools (role: ceo)

Memory Directory

hmem stores all memory files (.hmem SQLite databases) and its configuration (hmem.config.json) in a single directory. The location depends on how you install:

The hmem init installer asks which mode you prefer and creates the directory automatically.

Directory structure

~/.hmem/                     # System-wide memory directory
  memory.hmem                # Default agent memory (when no HMEM_AGENT_ID is set)
  SIGURD.hmem                # Named agent memory (HMEM_AGENT_ID=SIGURD)
  hmem.config.json           # Configuration file
  audit_state.json           # Curator state (optional)

The MCP configuration files are written to each tool's own config directory — not into ~/.hmem/:

Environment Variables

Configuration (hmem.config.json)

Place an optional hmem.config.json in your HMEM_PROJECT_DIR to tune behavior. All keys are optional — missing keys fall back to defaults.

{
  "maxL1Chars": 120,
  "maxLnChars": 50000,
  "maxDepth": 5,
  "accessCountTopN": 5,
  "recentDepthTiers": [
    { "count": 10, "depth": 2 },
    { "count": 3,  "depth": 3 }
  ],
  "prefixes": {
    "R": "Research"
  }
}

Memory prefixes

The default prefixes cover most use cases:

To add your own, add entries to the "prefixes" key in hmem.config.json. Custom prefixes are merged with the defaults — you don't need to repeat the built-in ones.

Favorites

Any entry can be marked as a favorite — regardless of its prefix category. Favorites always appear with their L2 detail in bulk reads, marked with [♥].

write_memory(prefix="D", content="...", favorite=true)     # set at creation
update_memory(id="D0010", content="...", favorite=true)    # set on existing entry
update_memory(id="D0010", content="...", favorite=false)   # clear the flag

Use favorites for reference info you need to see every session — key decisions, API endpoints, frequently consulted patterns. Use sparingly: if everything is a favorite, nothing is.

Access-count auto-promotion (accessCountTopN)

The top-N most-accessed entries are automatically promoted to L2 depth in bulk reads, marked with [★]. This creates "organic favorites" — entries that proved important in practice rise to the surface automatically.

{ "accessCountTopN": 5 }

Set to 0 to disable. Default: 5.

Token-efficient bulk reads

In a default read_memory() call, each entry shows only its most recently added L2 child (with that child's timestamp). A +N more hint indicates when additional L2 nodes exist. This keeps the bulk output compact while remaining discoverable.

To see all children of an entry, use read_memory(id="P0005").

Effective-date sorting

Entries are sorted by effective_date — the most recent timestamp across the entry and all its nodes. This means a project entry (P0005) that was first written months ago but had a new session note appended today will appear near the top of the listing, alongside truly recent entries.

Character limits

Two ways to set per-level character limits:

Option A — linear interpolation (recommended): set only the endpoints; all levels in between are computed automatically.

{ "maxL1Chars": 120, "maxLnChars": 50000 }

With 5 depth levels this yields: [120, 12780, 25440, 38120, 50000]

Option B — explicit per-level array: set each level individually. If fewer entries than maxDepth, the last value is repeated.

{ "maxCharsPerLevel": [120, 2500, 10000, 25000, 50000] }

Recency gradient (recentDepthTiers)

Controls how deep children are inlined for the most recent entries in a default read_memory() call. Each tier is { count, depth }: the count most recent entries get children inlined up to depth.

Tiers are cumulative — the highest applicable depth wins for each entry position.

"recentDepthTiers": [
  { "count": 3,  "depth": 3 },   // last 3 entries  → L1 + L2 + L3
  { "count": 10, "depth": 2 }    // last 10 entries → L1 + L2
]

Result:

This mirrors how human memory works: you remember today's events in full detail, last week's in outline, older ones only as headlines.

Set to [] to disable recency inlining (L1-only for all entries).

Backward compat: The old "recentChildrenCount": N key is still accepted and treated as [{ "count": N, "depth": 2 }].

Origin

hmem was developed out of necessity: working on a large AI project across multiple machines meant every new Claude Code session started blind. Agents redid work, lost decisions, and contradicted each other.

The solution was a memory protocol that works the way humans remember — broad strokes first, details on demand.

License

MIT