Kioku Lite

Kioku Lite — Agent Memory Skill

Kioku Lite is a long-term personal memory engine running fully local. Zero Docker, zero server. All data in SQLite. You (the agent) extract entities → call kg-index.

1. Language Handling

• Detect the user's language automatically. Always respond in the same language the user is writing in.
• Entity names: Extract AS-IS in the user's original language — do NOT translate.
• Entity types & relation types: ALWAYS use the predefined English labels (PERSON, EMOTION, TRIGGERED_BY, etc.).
• Evidence & saved text: Write in the user's original language.

Example (Vietnamese):
  "Hôm nay cãi nhau với sếp, cảm thấy rất căng thẳng"
  → entities: [{"name":"sếp","type":"PERSON"}, {"name":"Căng thẳng","type":"EMOTION"}]
  → relationships: [{"source":"Căng thẳng","rel_type":"TRIGGERED_BY","target":"Cãi nhau","evidence":"cãi nhau với sếp, cảm thấy rất căng thẳng"}]

Example (English):
  "Had an argument with my boss, feeling very stressed"
  → entities: [{"name":"boss","type":"PERSON"}, {"name":"Stress","type":"EMOTION"}]
  → relationships: [{"source":"Stress","rel_type":"TRIGGERED_BY","target":"Argument with boss","evidence":"had an argument with my boss, feeling very stressed"}]

2. Installation & Setup

Check if installed

kioku-lite --help

Install (if not found)

pipx install "kioku-lite[cli]"        # Global — recommended
# OR: pip install "kioku-lite[cli]"   # Inside a venv

(Optional) Pre-download embedding model (~1.1GB)

kioku-lite setup

Inject SKILL.md for agent discovery

kioku-lite init --global   # Global: works in ALL projects
# OR: kioku-lite init      # Per-project: creates ./AGENTS.md + .agents/skills/

3. Session Start — Run EVERY Session

Step A — Check active profile:

kioku-lite users

Step B — Ask user which profile to use (if multiple exist), then activate:

kioku-lite users --use <profile_name>
# To create a new profile: kioku-lite users --create <name>

Step C — Load context:

kioku-lite search "profile background goals recent" --limit 10

users --use only needs to run once per session. All subsequent commands use it automatically.

4. Command Reference

Write Commands

Query Commands

Dedup & Entity Management

Consolidation & Analysis

5. save — Store a Memory

kioku-lite save "TEXT" --mood MOOD --tags "tag1,tag2" --event-time "YYYY-MM-DD"

Output: JSON with content_hash → use immediately with kg-index.

Rules:

• ✅ Preserve full original text — do not summarize or paraphrase
• ✅ --event-time is REQUIRED whenever the event is not today. Search filters use this field. Parse relative dates: "hồi tháng 3 năm ngoái" → 2025-03-01, "yesterday" → yesterday's date, "năm 2019" → 2019-01-01. Omit only if the event is happening today or timing is truly unclear.
• ✅ Use the same --event-time on both save AND kg-index for consistency.
• ✅ Mood values: happy | sad | excited | anxious | grateful | proud | reflective | neutral | work | curious
• ❌ Do not add editorial comments — save raw information

Entry Splitting Strategy

SPLIT into multiple entries if ANY of these are true:

• ≥3 distinct topics (e.g. career + family + hobbies)
• ≥10 entities would be needed in kg-index
• ≥2 time phases with different contexts (e.g. "worked in Japan 5 years, then moved back")
• >300 words AND covers multiple topics

Keep as 1 entry if ALL of these are true:

• Single topic with single emotional arc
• <5 entities total
• Single time point or tight narrative

How to split: Group by phase → topic → emotion. Each entry should have 5–8 focused entities.

❌ BAD: 1 mega-entry (1500 words, 28 entities, 14 topics)
   → KG becomes tangled, search returns noise

✅ GOOD: 14 focused entries (~100-150 words, 5-8 entities each)
   → Clean KG, each entry is independently searchable
   → Use relationships to LINK entries across topics

6. kg-index — Index Entities & Relationships

After every save, you must: (1) disambiguate, (2) extract, (3) index.

Step 1 — Disambiguate: check existing entities

kioku-lite entities --limit 50

Compare extracted names against the returned list. Reuse existing canonical names instead of creating duplicates:

• If "Phúc" exists with 12 mentions, use "Phúc" — not "anh Phúc" or "Nguyễn Trọng Phúc"
• If "TechBase Vietnam" exists, use it — not "TBV" or "công ty"
• For true aliases, register them: kioku-lite kg-alias "Phúc" --aliases '["anh Phúc","Nguyễn Trọng Phúc"]'

Step 2 — Extract entities & relationships from the saved text

Entity types (generic): PERSON | PROJECT | PLACE | TOOL | CONCEPT | ORGANIZATION | EVENT

Relationship types (generic): KNOWS | WORKS_ON | WORKS_AT | CONTRIBUTED_TO | USED_BY | LOCATED_AT | INVOLVES | MENTIONS

Profile-specific types: If a persona profile (companion/mentor) is active, use the entity & relationship types from that profile's SKILL.md INSTEAD of the generic ones above.

Extraction rules:

• ✅ Use short, canonical name form: "Alice" not "my friend Alice"
• ✅ Prefer proper names over generic labels: "Phong" not "Con trai", "Sato" not "manager". Generic labels make recall fail.
• ✅ Entity names in the user's original language — do NOT translate
• ✅ evidence = exact quote from the saved text that supports the relationship
• ❌ Skip generic words: "I", "we", "they", "team", "everyone"
• ❌ Only add relationships explicitly stated in the text — do NOT infer
• ✅ No specific entities → skip kg-index entirely

Step 3 — Call kg-index with --event-time

kioku-lite kg-index <content_hash> \
  --entities '[{"name":"Alice","type":"PERSON"},{"name":"Project X","type":"PROJECT"}]' \
  --relationships '[
    {"source":"Alice","rel_type":"WORKS_ON","target":"Project X","evidence":"had a meeting with Alice about Project X"}
  ]' \
  --event-time "2024-06-15"

--event-time is critical for temporal accuracy. It sets the date on graph edges.

Parse relative time expressions to YYYY-MM-DD relative to today's date:

7. search — Enriched Search Workflow

Never call search with the raw user query. Always enrich first.

Step 1 — Analyze intent and enrich

Step 2 — Extract entities and build enriched query

🚨 ALWAYS pass --entities when the query mentions or implies specific people, places, or topics. This activates the graph search backend — without it, only vector+BM25 are used and relationship-based results are lost.

How: Identify entity names from the user's question, match them against known entities (kioku-lite entities), and pass as --entities.

--entities rules (Task 2H):

• ✅ Pass 2–3 specific entities relevant to the topic (other people, places, organizations)
• ❌ Do NOT add the user's own name — it connects to everything, dilutes results, and is auto-excluded by the engine
• ❌ Do NOT pass a single entity — use recall "Entity" instead for a more focused graph traversal

# Multi-topic: pass the relevant NON-user entities
kioku-lite search "công việc dạo này" --entities "Techbase,Brain,Sato" --limit 15

# ❌ Wrong: adding user name adds noise
kioku-lite search "..." --entities "Phúc,Techbase,Brain,Sato"

# Single entity → recall, not search
kioku-lite recall "Alice" --hops 2 --limit 15

# Connection between two entities
kioku-lite connect "Alice" "Bob"

# Temporal slice
kioku-lite search "events" --from 2026-02-01 --to 2026-02-28

# Recent timeline
kioku-lite timeline --limit 20

Step 3 — Interpret results

• Results contain content_hash — can be used for additional kg-index if needed
• 0 results → be honest, don't invent memories
• Low confidence (score < 0.02) → say "possibly related, but not certain"

8. Full Workflow Example

User: "Năm 2019, tôi quyết định quay lại đọc sách nghiêm túc."

─── Step 1: Save ───
kioku-lite save "Năm 2019, tôi quyết định quay lại đọc sách nghiêm túc." \
  --mood reflective --event-time 2019-01-01
→ {"content_hash": "a1b2c3..."}

─── Step 2: Disambiguate ───
kioku-lite entities --limit 50
→ Check: "Phúc" exists (12 mentions) — reuse it

─── Step 3: Extract & Index ───
kioku-lite kg-index a1b2c3 \
  --entities '[{"name":"Phúc","type":"PERSON"},{"name":"Đọc sách","type":"COPING_MECHANISM"},{"name":"Quyết định quay lại đọc sách","type":"LIFE_EVENT"}]' \
  --relationships '[
    {"source":"Phúc","rel_type":"TRIGGERED_BY","target":"Quyết định quay lại đọc sách","evidence":"quyết định quay lại đọc sách nghiêm túc"},
    {"source":"Quyết định quay lại đọc sách","rel_type":"REDUCED_BY","target":"Đọc sách","evidence":"quay lại đọc sách nghiêm túc"}
  ]' \
  --event-time 2019-01-01

8A. Temporal Facts Workflow (NEW)

Facts age. When a user's status changes (job, relationship, location), mark the old fact as superseded:

When to invalidate

• Job change: "I left LINE" → mark WORKS_AT "LINE" with --date 2026-03-31
• Relationship: "We broke up" → mark relevant edges as obsolete
• Project completion: "We shipped it" → mark WORKS_ON if active phase ended
• Resolved issue: "Finally fixed the bug" → mark BLOCKED_BY relationship

Command

kioku-lite kg-invalidate --source "Phuc" --target "LINE" --rel-type WORKS_AT \
  --date 2026-03-31 --reason "Changed jobs to Techbase"

Querying with temporal context

kioku-lite search "where does Phuc work?" 
# Returns current job only (superseded edges excluded)

kioku-lite search "where does Phuc work?" --include-historical
# Returns ALL jobs (including LINE)

Rule: Always set --date to when the fact became obsolete, not today. This preserves accurate historical context.

8B. Consolidation Workflow (NEW)

Periodic maintenance: detect stale edges, old memories, and duplicates. Agent reviews and acts:

Step 1 — Run consolidation

kioku-lite consolidate --half-life 90 --older-than 30

Output: JSON report with 3 sections:

{
  "decay": [
    {"source":"Phuc","target":"Techbase","current_weight":0.25,"reason":"Not reinforced 120 days"}
  ],
  "merge_suggestions": [
    {"entity1":"Phuc","entity2":"Phúc","sim":0.99}
  ],
  "stale_memories": [
    {"hash":"abc123...","content":"...","last_updated":"2025-06-01"}
  ]
}

Step 2 — Agent reviews and acts

On decayed edges:

• If still relevant → reinforce via new kg-index with same relationship
• If obsolete → kg-invalidate to supersede

On merge suggestions:

• Verify the pairs → merge "Phuc" "Phúc" to consolidate
• Or dedup-scan --auto to merge all qualifying pairs at once

On stale memories:

• Summarize old entries into weekly/monthly recap
• Save recap with save "Weekly summary: ..." --event-time 2025-12-31
• Delete originals if appropriate

Step 3 — Cleanup (optional)

# Auto-merge all detected duplicates
kioku-lite consolidate --auto-merge

# Or scan then auto-merge separately
kioku-lite dedup-scan --auto

Frequency: Run monthly or quarterly depending on memory growth (≥100 new edges/month).

8C. Entity Deduplication (NEW)

Agents may extract "Phuc" and "Phúc" as separate entities. Auto-dedup finds and merges them:

Scan for duplicates

kioku-lite dedup-scan
# Output: candidates with similarity scores

Automatic merge (recommended)

kioku-lite dedup-scan --auto
# Merges pairs where vector_sim ≥ 0.98 AND jaro_winkler ≥ 0.85

Manual merge (when auto is too strict)

kioku-lite merge "Phuc" "Phúc"
# source "Phuc" absorbed into target "Phúc"
# All edges re-pointed, "Phuc" added as alias

Rule: Dual-threshold (vector + name) prevents false merges. If merge fails, check name similarity manually.

8D. Cluster Analysis (NEW)

KGs fragment into communities (friend groups, work teams, interest clusters). View structure:

List all clusters

kioku-lite clusters
# Output: [{"label":"PERSON","entity_count":42}, {"label":"PROJECT","entity_count":8}, ...]

Inspect a cluster

kioku-lite cluster PERSON
# Shows all entities + connected memories in PERSON cluster

Use case

• Understand your KG structure (how many friend groups? work domains?)
• During consolidation, merge duplicates within clusters (more likely to be true duplicates)
• Identify isolated clusters that may need integration

12. Decision Tree

Session start?
└─ users → search to load context

User shares info / "remember this":
└─ Check splitting criteria → save (split if needed)
   → entities (disambiguate) → extract → kg-index each
   
User reports life change (job, breakup, move):
└─ kg-invalidate to mark old facts as obsolete

User asks a question:
└─ ENRICH query → search / recall / connect
   └─ Use --entities for multi-hop PPR search
   └─ Use --include-historical to include superseded facts

Periodic maintenance (monthly):
└─ consolidate → review decay/merge/stale → act

"What happened on [date]?":
└─ search --from DATE --to DATE

"Tell me about X":
└─ recall "X" --hops 2

"How are X and Y related?":
└─ connect "X" "Y"

Critical rules:

• 🚫 Never invent memories. 0 results → be honest.
• ✅ Always save when user shares valuable information.
• ✅ Always kg-index after save if entities are present (with --event-time!).
• ✅ Always disambiguate against existing entities before indexing.
• ✅ Always kg-invalidate when facts become obsolete (job change, breakup, etc.).
• ✅ Run consolidate monthly to surface stale edges and duplicates.
• ✅ Enrich queries — replace pronouns with real entity names.

13. Config & Data Locations

~/.kioku-lite/
└── users/<user_id>/        ← default user_id = "personal"
    ├── memory/             # Markdown backup (source of truth)
    └── data/
        └── kioku.db        # SQLite: FTS5 + sqlite-vec + KG tables + temporal + merge_log

Optional ~/.kioku-lite/config.env — only needed to change embedding provider or default user_id. Per-project override: echo "KIOKU_LITE_USER_ID=project-x" > .env

14. Troubleshooting