Keepsake — Memory Plugin for Hermes Agent

The Keepsake system automatically retrieves relevant memories and injects them into the conversation context for each dialogue.

User: "How did we set up that React project structure last time?"
                      ↓
         Keepsake System     ← Redis + RediSearch
                      ↓
    ┌─────────────────────────────────────┐
    │  [1] User prefers TypeScript + Vite    │
    │  [2] Previous projects used pinia state management │
    │  [3] Backend suggested using .NET 10 implementation       │
    └─────────────────────────────────────┘
                      ↓
         Model directly uses memories to answer

Features

• Full Entry Storage — stores complete text as-is, no semantic splitting
• BM25 Full-Text Search — works out of the box with no external API
• Optional Vector Search — KNN via RediSearch (OpenAI / DashScope embedder)
• Time Decay — newer entries rank higher (60-day half-life configurable)
• Sentiment Weighting — emotional entries get priority
• User Feedback — mark entries useful/useless to improve ranking
• Hot Topic Boost — frequently discussed topics rank higher
• Entity Extraction — auto-tags entries with entities (people, places, crypto tickers, domain terms) at store time; searched alongside content text for higher recall
• Entity Co-occurrence — auto-track which entities appear together, expand search to co-occurring entities for associative recall ("Python" → also finds entries mentioning "Django")
• Domain Dictionary — jieba user dictionary auto-generated from corpus + synonym table, loaded on /new for better Chinese tokenization
• Workflow Lock — set keepsake:workflow_lock in Redis to globally disable memory retrieval (e.g. during automated workflows)
• Skip Patterns — define skip lists (via file) to avoid searching on trivial queries like "ok", "got it"
• On-Demand Storage — only memory(action='add') stores data; no automatic per-turn archiving
• Search-Time Expiry — invalid_at field in index: set a timestamp and the entry is filtered out at search time (no data loss, can be reverted)
• Auto Maintenance — consolidation (keyword clustering + LLM summarization) + selective forgetting (multi-dimension low-value detection) run every 2h to keep storage tidy
• Auto-Registered Cron Jobs — when used as a Hermes plugin, three cron jobs (memory maintenance every 2h, deduplication every 1h, synonym discovery every 8h) are automatically registered on plugin initialization — zero manual setup
• Hermes Plugin Wrapper — ready-to-use hermes-plugin/ directory with plugin.yaml and __init__.py for drop-in installation

Design Philosophy: Clean Memory for LLMs

Keepsake stores full, self-contained entries — not entryed conversation snippets. The key insight is that LLMs need complete context to make use of stored information. A entry like "prefers TypeScript + Vite" without its surrounding context is useless; the full entry "User prefers TypeScript + Vite for frontend projects" is immediately actionable.

Mechanism	Implementation
Complete Context	Stores full text entries, no splitting
Forgetting Curve	Time decay (60-day half-life) — old memories fade naturally
Emotion Deepens Memory	Emotional weight boost — intense moments stick
Repetition Reinforces	Attention tracking + hot topic scoring
Use It or Lose It	Feedback reinforcement (keepsake_feedback)
Association & Analogy	Synonym discovery (Jaccard co-occurrence statistics) — "deploy" ↔ "release"
Entity Association	Entity co-occurrence tracking — entries mentioning "BTC" also recall "halving" without being synonyms
Entity Tagging	Like the brain tagging memories with people/places/things — auto-extracted entities searched alongside content
On-Demand Storage	No automatic archiving; only saves when explicitly told to (memory tool)
Sleep Consolidation	Background maintenance every 2h: keyword-based clustering + LLM summarization
Context Isolation	agent_id tagging — different identities, separate memories
Fuzzy but Enough	BM25 full-text search — doesn't need an exact match to recall

No vector database. No embedding API calls. No LLM inference for memory operations. Just pure statistical methods running on Redis + RediSearch — the same techniques the brain uses: frequency, recency, emotional salience, association, and feedback.

Requirements

• Python 3.10+
• Hermes Agent 0.12+ — provides MemoryProvider interface
• Redis 7+ — with RediSearch module (v2.6+)
• jieba — Chinese tokenization (auto-installed)
• Embedding API (optional) — OpenAI / DashScope / any compatible /v1/embeddings service

Installation

pip install keepsake

Or install directly from GitHub:

pip install git+https://github.com/j-zly/keepsake.git

Configuration

Configuration precedence (high to low): Environment variables > JSON config file > config.yaml inline > defaults

1. Configuration Methods

There are three ways to configure Keepsake, listed in order of priority:

1. Environment Variables (Highest precedence)
   Set environment variables like KEEPSAKE_REDIS_HOST, KEEPSAKE_REDIS_PASSWORD, etc.

2. JSON Config File (~/.config/keepsake/config.json)
   A complete JSON configuration file for all settings.

3. Code Defaults (Lowest precedence)
   Default values defined in the code.

2. Complete Configuration Example

Here's a comprehensive example of the configuration file ~/.config/keepsake/config.json with all available options:

{
  // Redis connection
  "redis_host": "127.0.0.1",
  "redis_port": 6379,
  "redis_password": "",
  
  // Search settings
  "top_k": 5,
  "candidate_k": 10,
  "bm25_limit": 10,
  "tag_filter": "",

  // Skip patterns
  "skip_min_length": 2,
  "skip_patterns_file": "~/.config/keepsake/skip_patterns.txt",

  // Time decay
  "decay_half_days": 60,
  "hot_topic_decay_half_days": 30,
  
  // Ranking weights
  "sentiment_boost_positive": 1.5,
  "sentiment_boost_negative": 1.3,
  "feedback_positive_boost": 1.3,
  "feedback_negative_penalty": 0.5,
  "hot_topic_boost": 1.2,
  
  // Attention
  "attention_boost_max": 1.5,
  "attention_base_increment": 2.0,
  "attention_emotion_factor": 1.5,
  
  // Embedding (optional)
  "embedder": {
    "provider": "dashscope",
    "api_key": "sk-xxx",
    "base_url": "https://dashscope.aliyuncs.com/compatible-mode/v1",
    "model": "text-embedding-v2"
  },
  
  // Auto maintenance
  "consolidate_min_group": 2,
  "consolidate_max_age_hours": 72,
  "forget_max_age_days": 30,
  "forget_dry_run": false,

  // Agent isolation
  "agent_id": "main-brain",
  "is_primary": true,

  // Synonym discovery
  "synonym_min_word_freq": 10,
  "synonym_jaccard_threshold": 0.5,
  "synonym_min_co_occurrence": 3,

  // Entity co-occurrence
  "entity_cooc_top_n": 3,
  "entity_cooc_min_count": 2,
  
  // Emotion intensity factor
  "emotion_intensity_factor": 0.4
}

Note: Redis password compatibility: leave empty for no authentication, or provide password to automatically send AUTH command.

3. Environment Variables Reference

Environment Variable	Corresponding Config Item	Description
KEEPSAKE_REDIS_HOST	redis_host	Redis server host
KEEPSAKE_REDIS_PORT	redis_port	Redis server port
KEEPSAKE_REDIS_PASSWORD	redis_password	Redis password for authentication
KEEPSAKE_TOP_K	top_k	Number of final entries returned
KEEPSAKE_CANDIDATE_K	candidate_k	Candidate entries count (for KNN)
KEEPSAKE_BM25_LIMIT	bm25_limit	BM25 search candidate count
KEEPSAKE_TAG_FILTER	tag_filter	Tag filtering (comma-separated)
KEEPSAKE_DECAY_HALF_DAYS	decay_half_days	Time decay half-life (days)
KEEPSAKE_HOT_TOPIC_DECAY_HALF_DAYS	hot_topic_decay_half_days	Hot topic time decay half-life (days)
KEEPSAKE_EMBED_CACHE_TTL	embed_cache_ttl	Embedding cache TTL (seconds)
KEEPSAKE_EMBEDDER	embedder.provider	Embedding provider (openai, dashscope)
KEEPSAKE_EMBEDDER_URL	embedder.base_url	Embedding API endpoint
KEEPSAKE_EMBEDDER_MODEL	embedder.model	Embedding model name
KEEPSAKE_CONSOLIDATE_MIN_GROUP	consolidate_min_group	Minimum entries to trigger consolidation
KEEPSAKE_CONSOLIDATE_MAX_AGE_HOURS	consolidate_max_age_hours	Minimum age (hours) before entries can be consolidated
KEEPSAKE_FORGET_MAX_AGE_DAYS	forget_max_age_days	Number of days before entries might be forgotten
KEEPSAKE_FORGET_DRY_RUN	forget_dry_run	Safe mode: true = count only, false = actually delete
KEEPSAKE_EMOTION_INTENSITY_FACTOR	emotion_intensity_factor	Emotion intensity → weight coefficient (0=disabled, 1=max)

Note: Redis password is compatible with empty value (no auth) or password provided for AUTH command.
Note: Changes to config.json take effect immediately without restarting (just send /new).

4. Create Redis Index (first-time usage)

The code will auto-create (ensure_index()), or execute manually:

redis-cli FT.CREATE idx:memories ON HASH PREFIX 1 "memory:frag:" SCHEMA \
    content TEXT WEIGHT 1 \
    tags TAG SEPARATOR "," \
    category TAG SEPARATOR "," \
    source TEXT WEIGHT 1 \
    created TEXT WEIGHT 0 \
    entry_type TAG SEPARATOR "," \
    invalid_at TAG SEPARATOR "," \
    entities TAG SEPARATOR "," \
    embed_bin VECTOR FLAT 6 TYPE FLOAT32 DIM 1536 DISTANCE_METRIC COSINE

Dimension (DIM) is dynamically adjusted based on the embedding model used, default 1536. For Docker: docker run -d --name redis-stack -p 6379:6379 redis/redis-stack:latest

5. Hermes Configuration

Enable in ~/.hermes/config.yaml:

memory:
  provider: keepsake

If embedder is not configured, only BM25 full-text search mode will be used.

Also supports environment variable configuration (highest precedence):

export KEEPSAKE_REDIS_HOST=127.0.0.1
export KEEPSAKE_REDIS_PORT=6379
export KEEPSAKE_TOP_K=5
export KEEPSAKE_EMBEDDER=dashscope
export KEEPSAKE_EMBEDDER_MODEL=text-embedding-v2
export OPENAI_API_KEY=sk-xxx        # embedder API key

6. Workflow Lock

Temporarily disable memory retrieval during automated workflows (like batch processing):

# Lock (3600s TTL)
redis-cli SET keepsake:workflow_lock 1 EX 3600

# Unlock
redis-cli DEL keepsake:workflow_lock

7. Skip Patterns File

Create a file (one pattern per line, # for comments):

# ~/.config/keepsake/skip_patterns.txt
好的
嗯
对
是
哦
可以
没错
ok
okay
yes
yeah

Then reference it in config.json:

{
  "skip_min_length": 2,
  "skip_patterns_file": "~/.config/keepsake/skip_patterns.txt"
}

8. Restart Gateway

# For CLI mode, restart session is sufficient
# For Gateway mode, restart the process

Configuration Reference

Config Item	Environment Variable	Default Value	Description
redis_host	KEEPSAKE_REDIS_HOST	127.0.0.1	Redis address
redis_port	KEEPSAKE_REDIS_PORT	6379	Redis port
top_k	KEEPSAKE_TOP_K	5	Number of final entries returned
candidate_k	KEEPSAKE_CANDIDATE_K	10	Candidate entries count (for KNN)
tag_filter	KEEPSAKE_TAG_FILTER	""	Tag filtering (comma-separated)
bm25_limit	KEEPSAKE_BM25_LIMIT	10	BM25 search candidate count
decay_half_days	KEEPSAKE_DECAY_HALF_DAYS	60	Time decay half-life (days)
embed_cache_ttl	KEEPSAKE_EMBED_CACHE_TTL	3600	Embedding cache TTL (seconds)
sentiment_boost_positive	—	1.5	Positive entry weight multiplier
sentiment_boost_negative	—	1.3	Negative entry weight multiplier
feedback_positive_boost	—	1.3	Positive feedback bonus weight
feedback_negative_penalty	—	0.5	Negative feedback penalty coefficient
hot_topic_boost	—	1.2	Hot topic weighting multiplier
embedder.provider	KEEPSAKE_EMBEDDER	openai	openai / dashscope
embedder.api_key	OPENAI_API_KEY	—	Embedding API key
embedder.base_url	KEEPSAKE_EMBEDDER_URL	https://api.openai.com/v1	API endpoint
embedder.model	KEEPSAKE_EMBEDDER_MODEL	text-embedding-3-small	Embedding model name
consolidate_min_group	—	2	Minimum entries to trigger consolidation
consolidate_max_age_hours	—	72	Minimum age (hours) before consolidation
forget_max_age_days	—	30	Max age (days) before deletion
forget_dry_run	—	true	Safe mode: true = count only, false = delete
agent_id	—	""	Agent identity tag for isolation (e.g. "main-brain")
is_primary	—	false	true = sees all entries; false = only tagged ones
hot_topic_decay_half_days	—	30	Hot topic time decay half-life (days)
emotion_intensity_factor	—	0.4	Emotion intensity → weight coefficient
skip_min_length	—	2	Minimum query length to trigger search
skip_patterns_file	—	""	Path to skip patterns file
attention_boost_max	—	1.5	Max attention weighting value
attention_base_increment	—	2.0	Base attention increment per mention
attention_emotion_factor	—	1.5	Emotion amplification for attention
synonym_min_word_freq	—	10	Min frequency for synonym candidate
synonym_jaccard_threshold	—	0.5	Jaccard threshold for synonym detection
synonym_min_co_occurrence	—	3	Min co-occurrence for synonym detection
entity_cooc_top_n	—	3	Number of co-occurring entities to expand search
entity_cooc_min_count	—	2	Min co-occurrence for entity association

sentiment_*, feedback_*, hot_topic_* and other ranking weight parameters currently only support configuration through JSON config file, not environment variables. Set to 1.0 to disable the effect of that dimension.

Embedding Models and Dimensions

Model	Dimensions
OpenAI text-embedding-3-small	1536
OpenAI text-embedding-3-large	3072
OpenAI text-embedding-ada-002	1536
DashScope text-embedding-v2	1536
DashScope text-embedding-v3	1024

Dimensions are automatically detected, switching models doesn't require reconfiguration.

Synonym Table

Stored in Redis Hash keepsake:synonyms, expanded at search time to improve recall:

redis-cli HSET keepsake:synonyms setup '["install","configure","deploy","setup"]'
redis-cli HSET keepsake:synonyms fix '["fix","modify","correct","repair","solve"]'

Verification

Check logs after startup:

Memory provider 'keepsake' registered (0 tools)
keepsake: connected (session=xxx, top_k=5, tag_filter=(none))
keepsake: BM25-only mode (no embedder configured)
keepsake: auto-registered cron job 'memory-maintenance'
keepsake: auto-registered cron job 'synonym-discovery-daily'
keepsake: auto-registered cron job '记忆去重'

Project Structure

keepsake/
├── src/keepsake/         # Python package — the memory provider
├── hermes-plugin/        # Hermes plugin wrapper (drop-in for ~/.hermes/plugins/)
│   ├── plugin.yaml
│   └── __init__.py
├── cron/                 # Cron wrapper scripts for scheduled tasks
│   ├── memory-maintenance.py   # Every 2h — consolidation + forgetting
│   ├── dedup-memory.sh         # Every 1h — deduplication
│   └── discover-synonyms.py    # Every 8h — synonym auto-discovery
├── scripts/              # Standalone utility scripts (dev/test)
├── README.md
└── pyproject.toml

The three cron jobs in cron/ are auto-registered when the keepsake plugin initializes (on /new or gateway restart). No manual hermes cron create needed.

Architecture

┌────────────────────────────────────────────────────────┐
│                    User sends message                  │
└──────────────────┬─────────────────────────────────────┘
                   │
         ┌─────────▼─────────┐
         │   prefetch()       │  ← Automatically triggered on every user message
         │   ↓                │
         │  Workflow Lock?    │  ← Checks keepsake:workflow_lock
         │   ↓                │
         │  Skip patterns?    │  ← Length / exact match against skip list
         │   ↓                │
         │  BM25 Full-Text Search │  ← Default, zero cost, searches full entries
         │  (KNN Vector search) │  ← Optional (needs embedder)
         │  Entity co-occurrence │  ← Expand query with co-occurring entities
         │   ↓                │
         │  Six-dimensional Re-ranking │  ← Similarity × Time decay
         │                    │    × Emotion × Feedback × Hot Topic × Attention
         │   ↓                │
         │  Top N Injected into Context   │  ← Full entries returned as-is
         └─────────┬─────────┘
                   │
         ┌─────────▼─────────┐
         │   Model Response   │  ← Entries used directly (complete text)
         └───────────────────┘
                   │
         ┌─────────▼─────────┐
         │   on_memory_write()│  ← Only on memory(action='add')
         │   Stores Full Text │  ← Complete entry, no splitting
         │   Entity Extraction│  ← jieba + regex → entities TAG field
         │   Entity Co-occur. │  ← Track entity pairs in ZSET
         │   Attention Track  │  ← Extract keywords, increase attention score
         │   ↓                │
         │   Stored in Redis  │  ← Available for next retrieval as full text
         └───────────────────┘
                   │
         ┌─────────▼─────────┐
         │   [cron] Every 2h     │  ← Background maintenance
         │   ① Multi-level Consolidation  │  ← Same topic → keyword clustering → LLM → level+1
         │   ② Selective Forgetting  │  ← Age>30d + no feedback + low emotion + low attention → delete
         └───────────────────┘

License

MIT