remembering

name	remembering
description	Advanced memory operations reference. Basic patterns (profile loading, simple recall/remember) are in project instructions. Consult this skill for background writes, memory versioning, complex queries, edge cases, session scoping, retention management, type-safe results, proactive memory hints, GitHub access detection, and ops priority ordering.
metadata	{"version":"3.7.0"}

⚠️ IMPORTANT FOR CLAUDE CODE AGENTS Before working with this skill, read CLAUDE.md in this directory. It contains critical development context, import patterns, and instructs you to use Muninn to track work on Muninn (meta-usage pattern).

Remembering - Advanced Operations

Basic patterns are in project instructions. This skill covers advanced features and edge cases.

Two-Table Architecture

Table	Purpose	Growth
config	Stable operational state (profile + ops + journal)	Small, mostly static
memories	Timestamped observations	Unbounded

Config loads fast at startup. Memories are queried as needed.

Boot Sequence

Load context at conversation start to maintain continuity across sessions.

Optimized: Compressed Boot (Recommended)

Use boot() for fast startup (~150ms):

from remembering import boot
print(boot())

Output: Complete profile and ops values for full context at boot.

Performance:

• Execution: ~150ms (single HTTP request)
• Populates local cache for fast subsequent recall()

Boot Output: CAPABILITIES Section (v3.5.0)

Boot now includes a # CAPABILITIES section reporting:

GitHub Access:

• Detects gh CLI availability and authentication status
• Checks for GITHUB_TOKEN / GH_TOKEN environment variables
• Reports recommended method (gh-cli preferred when authenticated)
• Shows authenticated user when available

# CAPABILITIES

## GitHub Access
  Status: Available
  Methods: gh-cli, api-token
  Recommended: gh-cli
  gh user: oaustegard
  Usage: gh pr view, gh issue list, gh api repos/...

Utilities:

• Extracts utility-code memories to /home/claude/muninn_utils/
• Adds to Python path for direct import
• Lists available utilities with import syntax

## Utilities (2)
  from muninn_utils import my_helper
  from muninn_utils import another_util

Detecting GitHub Access Programmatically

from remembering import detect_github_access

github = detect_github_access()
if github['available']:
    print(f"Use {github['recommended']} for GitHub operations")
    if github['gh_cli'] and github['gh_cli']['authenticated']:
        print(f"Authenticated as: {github['gh_cli']['user']}")

Returns:

{
    'available': True,
    'methods': ['gh-cli', 'api-token'],
    'recommended': 'gh-cli',
    'gh_cli': {'path': '/usr/bin/gh', 'authenticated': True, 'user': 'username'},
    'api_token': True
}

Journal System

Temporal awareness via rolling journal entries in config. Inspired by Strix's journal.jsonl pattern.

from remembering import journal, journal_recent, journal_prune

# Record what happened this interaction
journal(
    topics=["project-x", "debugging"],
    user_stated="Will review PR tomorrow",
    my_intent="Investigating memory leak"
)

# Boot: load recent entries for context
for entry in journal_recent(10):
    print(f"[{entry['t'][:10]}] {entry.get('topics', [])}: {entry.get('my_intent', '')}")

# Maintenance: keep last 40 entries
pruned = journal_prune(keep=40)

Entry structure:

• t: ISO timestamp
• topics: array of tags (enables filtering at scale)
• user_stated: commitments/plans user verbalized
• my_intent: current goal/task

Key insight from Strix: "If you didn't write it down, you won't remember it next message."

Config Table

Key-value store for profile (behavioral), ops (operational), and journal (temporal) settings.

from remembering import config_get, config_set, config_delete, config_list, config_set_boot_load, config_set_priority, profile, ops

# Read
config_get("identity")                    # Single key
profile()                                  # All profile entries
ops()                                      # All ops entries
config_list()                              # Everything

# Write
config_set("new-key", "value", "profile")  # Category: 'profile', 'ops', or 'journal'
config_set("skill-foo", "usage notes", "ops")

# Write with constraints (new!)
config_set("bio", "Short bio here", "profile", char_limit=500)  # Enforce max length
config_set("core-rule", "Never modify this", "ops", read_only=True)  # Mark immutable

# Delete
config_delete("old-key")

Config constraints:

• char_limit: Enforces maximum character count on writes (raises ValueError if exceeded)
• read_only: Prevents modifications (raises ValueError on attempted updates)

Progressive Disclosure (v2.1.0)

Ops entries can be marked as boot-loaded (default) or reference-only to reduce boot() output size:

from remembering import config_set_boot_load, ops

# Mark entry as reference-only (won't load at boot)
config_set_boot_load('github-api-endpoints', False)
config_set_boot_load('container-limits', False)

# Mark entry as boot-loaded (loads at boot)
config_set_boot_load('storage-discipline', True)

# Query ops with filtering
boot_ops = ops()                          # Only boot-loaded entries (default)
all_ops = ops(include_reference=True)     # All entries (boot + reference)

How it works:

• boot() outputs only ops with boot_load=1 (reduces token usage at boot)
• Reference-only ops (boot_load=0) appear in a Reference Entries index at the end of boot output
• Reference entries remain fully accessible via config_get(key) when needed
• Ideal for: API documentation, container specs, rarely-triggered guidance

Example boot output:

=== OPS ===

## Core Boot & Behavior
storage-discipline:
[Full content here...]

## Reference Entries (load via config_get)
container-limits, github-api-endpoints, network-tools, recall-triggers

Priority-Based Ordering (v3.6.0)

Ops entries within each topic category are sorted by priority (descending). Critical entries appear first.

from remembering import config_set_priority

# Set priority for critical entries (higher = more important)
config_set_priority('storage-rules', 10)       # Critical - show first in category
config_set_priority('boot-behavior', 5)        # Elevated priority
config_set_priority('fly-command', 0)          # Normal priority (default)

# Priority scale:
# 0: Normal (default)
# 1-9: Elevated
# 10+: Critical

How it works:

• Entries with higher priority appear first within their topic category
• Equal-priority entries are sorted alphabetically by key
• Priority is stored in the priority column of the config table
• Does not require entry recreation—just call config_set_priority()

Dynamic Topic Categories (v3.6.0)

Topic categories can be loaded from config instead of being hardcoded:

from remembering import config_set, config_get
import json

# View current topic mapping
topics = json.loads(config_get('ops-topics') or '{}')

# Update topic mapping
new_topics = {
    'Core Boot & Behavior': ['boot-behavior', 'dev-workflow'],
    'Memory Operations': ['remembering-api', 'storage-rules'],
    'My Custom Category': ['my-key-1', 'my-key-2']
}
config_set('ops-topics', json.dumps(new_topics), 'ops')

How it works:

• boot() loads topics from config_get('ops-topics') if available
• Falls back to built-in defaults if config is missing or invalid
• Format: JSON object mapping topic name → list of ops keys
• Changes take effect on next boot() call

Memory Type System

Type is required on all write operations. Valid types:

Type	Use For
decision	Explicit choices: prefers X, always/never do Y
world	External facts: tasks, deadlines, project state
anomaly	Errors, bugs, unexpected behavior
experience	General observations, catch-all

Note: profile is no longer a memory type—use config_set(key, value, "profile") instead.

from remembering import TYPES  # {'decision', 'world', 'anomaly', 'experience'}

Priority System (v2.0.0)

Memories have a priority field that affects ranking in search results:

Priority	Value	Description
Background	-1	Low-value, can age out first
Normal	0	Default for new memories
Important	1	Boosted in ranking
Critical	2	Always surface, never auto-age

from remembering import remember, reprioritize

# Set priority at creation
remember("Critical security finding", "anomaly", tags=["security"], priority=2)

# Adjust priority later
reprioritize("memory-uuid", priority=1)  # Upgrade to important
reprioritize("memory-uuid", priority=-1)  # Demote to background

Ranking formula:

score = bm25_score * recency_weight * (1 + priority * 0.5)

Priority affects composite ranking score - higher priority memories surface more readily in search results.

Memory Consolidation (v3.3.0)

Biological memory consolidation pattern: memories that participate in active cognition consolidate more strongly.

from remembering import strengthen, weaken, recall

# Strengthen a memory (increment priority, max 2)
result = strengthen("memory-uuid", boost=1)
# Returns: {'memory_id': '...', 'old_priority': 0, 'new_priority': 1, 'changed': True}

# Weaken a memory (decrement priority, min -1)
result = weaken("memory-uuid", drop=1)
# Returns: {'memory_id': '...', 'old_priority': 1, 'new_priority': 0, 'changed': True}

# Auto-strengthen top results during recall (opt-in)
results = recall("important topic", auto_strengthen=True, n=10)
# Automatically strengthens top 3 results with priority < 2

Use cases:

• Consolidate memories that prove useful across conversations
• Implement spaced repetition patterns
• Automatically promote frequently accessed knowledge
• Simulate biological memory consolidation mechanisms

Notes:

• strengthen() caps at priority=2 (critical)
• weaken() floors at priority=-1 (background)
• auto_strengthen=True only affects top 3 results with priority < 2
• Returns dict with old/new priority and whether change occurred
• Replaced no-op placeholder functions from v2.0.0 with working implementations

Background Writes (Agentic Pattern)

v0.6.0: Unified API with sync parameter. Use remember(..., sync=False) for background writes:

from remembering import remember, flush

# Background writes (non-blocking, returns immediately)
remember("User's project uses Python 3.12 with FastAPI", "world", sync=False)
remember("Discovered: batch insert reduces latency 70%", "experience",
         tags=["optimization"], sync=False)

# Ensure all pending writes complete before conversation end
flush()  # Blocks until all background writes finish

Backwards compatibility: remember_bg() still works (deprecated, calls remember(..., sync=False)):

from remembering import remember_bg
remember_bg("Quick note", "world")  # Same as remember(..., sync=False)

When to use sync=False (background):

• Storing derived insights during active work
• Memory write shouldn't block response
• Agentic pattern where latency matters

When to use sync=True (blocking, default):

• User explicitly requests storage
• Need confirmation of write success
• Critical memories (handoffs, decisions)
• End of workflow when durability matters

⚠️ IMPORTANT - Cache Sync Guarantee:

• If you use sync=False for ANY writes in a conversation, you MUST call flush() before the conversation ends
• This ensures all background writes persist to the database before the ephemeral container is destroyed
• Single-user context: no concurrent write conflicts, all writes will succeed
• Prefer sync=True (default) for critical writes to guarantee immediate persistence

Memory Versioning (Patch/Snapshot)

Supersede without losing history:

from remembering import supersede

# User's preference evolved
original_id = "abc-123"
supersede(original_id, "User now prefers Python 3.12", "decision", conf=0.9)

Creates new memory with refs=[original_id]. Original preserved but not returned in default queries. Trace evolution via refs chain.

v3.3.0 Performance: supersede() now uses batched operations, reducing HTTP requests by 50% (single request instead of two).

Complex Queries

Multiple filters, custom confidence thresholds:

from remembering import recall

# High-confidence decisions only
decisions = recall(type="decision", conf=0.85, n=20)

# Recent anomalies for debugging context
bugs = recall(type="anomaly", n=5)

# Search with tag filter (any match)
tasks = recall("API", tags=["task"], n=15)

# Require ALL tags (tag_mode="all")
urgent_tasks = recall(tags=["task", "urgent"], tag_mode="all", n=10)

Date-Filtered Queries

Query memories by temporal range:

from remembering import recall_since, recall_between

# Get memories after a specific timestamp
recent = recall_since("2025-12-01T00:00:00Z", n=50)
recent_bugs = recall_since("2025-12-20T00:00:00Z", type="anomaly", tags=["critical"])

# Get memories within a time range
december = recall_between("2025-12-01T00:00:00Z", "2025-12-31T23:59:59Z", n=100)
sprint_mems = recall_between("2025-12-15T00:00:00Z", "2025-12-22T00:00:00Z",
                             type="decision", tags=["sprint-5"])

Use cases:

• Review decisions made during a project phase
• Analyze bugs discovered in a time window
• Track learning progress over specific periods
• Build time-based memory summaries

Notes:

• Timestamps are exclusive (use > and < not >= and <=)
• Supports all standard filters: search, type, tags, tag_mode
• Sorted by timestamp descending (newest first)
• Excludes soft-deleted and superseded memories

Therapy Helpers

Support for reflection and memory consolidation workflows:

from remembering import therapy_scope, therapy_session_count

# Get unprocessed memories since last therapy session
cutoff_time, unprocessed_memories = therapy_scope()
# cutoff_time: timestamp of last therapy session (or None if no sessions)
# unprocessed_memories: all memories created after that timestamp

# Count how many therapy sessions have been recorded
count = therapy_session_count()

Therapy session workflow:

1. Call therapy_scope() to get unprocessed memories
2. Analyze and consolidate memories (group patterns, extract insights)
3. Record therapy session completion:

   remember(f"Therapy Session #{count+1}: Consolidated {len(unprocessed)} memories...",
            "experience", tags=["therapy"])
   

Pattern detection example:

cutoff, mems = therapy_scope()
by_type = group_by_type(mems)  # See Analysis Helpers below

print(f"Since {cutoff}:")
print(f"  {len(by_type.get('decision', []))} decisions")
print(f"  {len(by_type.get('anomaly', []))} anomalies to investigate")

Analysis Helpers

Group and organize memories for pattern detection:

from remembering import group_by_type, group_by_tag

# Get memories and group by type
memories = recall(n=100)
by_type = group_by_type(memories)
# Returns: {"decision": [...], "world": [...], "anomaly": [...], "experience": [...]}

# Group by tags
by_tag = group_by_tag(memories)
# Returns: {"ui": [...], "bug": [...], "performance": [...], ...}
# Note: Memories with multiple tags appear under each tag

Use cases:

• Pattern detection: Find clusters of related memories
• Quality analysis: Identify over/under-represented memory types
• Tag hygiene: Discover inconsistent tagging patterns
• Therapy sessions: Organize unprocessed memories before consolidation

Example - Find overused tags:

mems = recall(n=200)
by_tag = group_by_tag(mems)
sorted_tags = sorted(by_tag.items(), key=lambda x: len(x[1]), reverse=True)
print("Top tags:")
for tag, tagged_mems in sorted_tags[:5]:
    print(f"  {tag}: {len(tagged_mems)} memories")

FTS5 Search with Porter Stemmer (v0.13.0)

Full-text search uses FTS5 with Porter stemmer for morphological variant matching:

from remembering import recall

# Searches match word variants automatically
# "running" matches "run", "runs", "runner"
# "beads" matches "bead"
results = recall("running performance")

# Query expansion fallback
# When FTS5 returns < threshold results, automatically extracts tags from
# partial results and searches for related memories
sparse_results = recall("rare term")  # Auto-expands if < 3 matches

# v3.7.0: Configurable expansion threshold
results = recall("term", expansion_threshold=5)  # Expand if < 5 results
results = recall("term", expansion_threshold=0)  # Disable expansion entirely

How it works:

• FTS5 tokenizer: porter unicode61 handles stemming
• BM25 ranking for relevance scoring
• Query expansion extracts tags from partial results when below threshold (default 3, configurable via expansion_threshold)
• Set expansion_threshold=0 to disable expansion entirely
• Composite ranking: BM25 × salience × recency × access patterns

Soft Delete

Remove without destroying data:

from remembering import forget

forget("memory-uuid")  # Sets deleted_at, excluded from queries

Memories remain in database for audit/recovery. Hard deletes require direct SQL.

Memory Quality Guidelines

Write complete, searchable summaries that standalone without conversation context:

✓ "User prefers direct answers with code examples over lengthy conceptual explanations"

✗ "User wants code" (lacks context, unsearchable)

✗ "User asked question" + "gave code" + "seemed happy" (fragmented, no synthesis)

Handoff Convention

Cross-environment work coordination with version tracking and automatic completion marking.

Creating Handoffs

From Claude.ai (web/mobile) - cannot persist file changes:

from remembering import remember

remember("""
HANDOFF: Implement user authentication

## Context
User wants OAuth2 + JWT authentication for the API.

## Files to Modify
- src/auth/oauth.py
- src/middleware/auth.py
- tests/test_auth.py

## Implementation Notes
- Use FastAPI OAuth2PasswordBearer
- JWT tokens with 24h expiry
- Refresh token support
...
""", "world", tags=["handoff", "pending", "auth"])

Important: Tag with ["handoff", "pending", ...] so it appears in handoff_pending() queries.

Handoff structure:

• Title: Brief summary of what needs to be done
• Context: Why this work is needed
• Files to Modify: Specific paths
• Implementation Notes: Code patterns, constraints, dependencies

Completing Handoffs

From Claude Code - streamlined workflow:

from remembering import handoff_pending, handoff_complete

# Get pending work (excludes completed handoffs)
pending = handoff_pending()
print(f"{len(pending)} pending handoff(s)")

for h in pending:
    print(f"[{h['created_at'][:10]}] {h['summary'][:80]}")

# Complete a handoff (automatically tags with version)
handoff_id = pending[0]['id']
handoff_complete(
    handoff_id,
    "COMPLETED: Implemented boot() function with batched queries...",
    # version auto-detected from VERSION file, or specify: "0.5.0"
)

What happens:

• Original handoff is superseded (won't appear in future handoff_pending() queries)
• Completion record created with tags ["handoff-completed", "v0.5.0"]
• Version tracked automatically from VERSION file
• Full history preserved via supersede() chain

Querying History

from remembering import recall

# See what was completed in a specific version
v050_work = recall(tags=["handoff-completed", "v0.5.0"])

# See all completion records
completed = recall(tags=["handoff-completed"], n=50)

Use when:

• Working in Claude.ai (web/mobile) without file write access
• Planning work that needs Claude Code execution
• Coordinating between environments
• Leaving detailed instructions for future sessions

Session Scoping (v3.2.0)

Filter memories by conversation or work session using session_id:

from remembering import remember, recall, set_session_id

# Set session for all subsequent remember() calls
set_session_id("project-alpha-sprint-1")
remember("Feature spec approved", "decision", tags=["project-alpha"])

# Query by session
alpha_memories = recall(session_id="project-alpha-sprint-1", n=50)

# Session ID defaults to MUNINN_SESSION_ID env var or 'default-session'
import os
os.environ['MUNINN_SESSION_ID'] = 'my-session'

Note: Session filtering bypasses cache (queries Turso directly). Cache support planned for future release.

Retrieval Observability (v3.2.0)

Monitor query performance and usage patterns:

from remembering import recall_stats, top_queries

# Get retrieval statistics
stats = recall_stats(limit=100)
print(f"Cache hit rate: {stats['cache_hit_rate']:.1%}")
print(f"Avg query time: {stats['avg_exec_time_ms']:.1f}ms")

# Find most common searches
for query_info in top_queries(n=10):
    print(f"{query_info['query']}: {query_info['count']} times")

Retention Management (v3.2.0)

Analyze memory distribution and prune old/low-priority memories:

from remembering import memory_histogram, prune_by_age, prune_by_priority

# Get memory distribution
hist = memory_histogram()
print(f"Total: {hist['total']}")
print(f"By type: {hist['by_type']}")
print(f"By priority: {hist['by_priority']}")
print(f"By age: {hist['by_age_days']}")

# Preview what would be deleted (dry run)
result = prune_by_age(older_than_days=90, priority_floor=0, dry_run=True)
print(f"Would delete {result['count']} memories")

# Actually delete old low-priority memories
result = prune_by_age(older_than_days=90, priority_floor=0, dry_run=False)

# Delete all background-priority memories
result = prune_by_priority(max_priority=-1, dry_run=False)

Export/Import for Portability

Backup or migrate Muninn state across environments:

from remembering import muninn_export, muninn_import
import json

# Export all state to JSON
state = muninn_export()
# Returns: {"version": "1.0", "exported_at": "...", "config": [...], "memories": [...]}

# Save to file
with open("muninn-backup.json", "w") as f:
    json.dump(state, f, indent=2)

# Import (merge with existing data)
with open("muninn-backup.json") as f:
    data = json.load(f)
stats = muninn_import(data, merge=True)
print(f"Imported {stats['config_count']} config, {stats['memory_count']} memories")

# Import (replace all - destructive!)
stats = muninn_import(data, merge=False)

Notes:

• merge=False deletes all existing data before import (use with caution!)
• Memory IDs are regenerated on import to avoid conflicts
• Returns stats dict with counts and any errors

Type-Safe Results (v3.4.0)

recall(), recall_since(), and recall_between() now return MemoryResult objects that validate field access immediately:

from remembering import recall, MemoryResult, VALID_FIELDS

# Recall returns MemoryResultList of MemoryResult objects
memories = recall("search term", n=10)

for m in memories:
    # Valid access - works fine
    print(m.summary)      # Attribute-style
    print(m['summary'])   # Dict-style
    print(m.get('summary', 'default'))  # get() with default

    # v3.7.0: Common aliases resolve transparently
    print(m.content)      # Works! Resolves to m.summary
    print(m['content'])   # Works! Resolves to m['summary']
    print(m.conf)         # Works! Resolves to m.confidence

    # Truly invalid fields still raise errors
    print(m.foo)          # AttributeError with list of valid fields

Parameter aliases (v3.7.0):

# limit= accepted as alias for n= in recall()
recall("search", limit=20)   # Same as recall("search", n=20)

Valid fields:

from remembering import VALID_FIELDS
# {'id', 'type', 't', 'summary', 'summary_preview', 'confidence', 'tags',
#  'refs', 'priority', 'session_id', 'created_at', 'updated_at', 'valid_from',
#  'access_count', 'last_accessed', 'bm25_score', 'composite_rank',
#  'composite_score', 'has_full', 'deleted_at'}

Transparent aliases (v3.7.0):

Alias	Resolves To
m.content	m.summary
m['text']	m['summary']
m.conf	m.confidence
m.timestamp	m.t
m.created	m.created_at

Normalized fields (v3.7.0):

All results now include summary_preview (first 100 chars) regardless of whether they came from cache or Turso. Fields like bm25_score and composite_rank are valid but only present on cache-sourced results.

Backward compatibility:

• MemoryResult supports all dict operations: in, len(), iteration, keys(), values(), items()
• Use m.to_dict() to convert back to plain dict when needed
• Use raw=True parameter to get plain dicts: recall("term", raw=True)

Proactive Memory Hints (v3.4.0)

recall_hints() scans context for terms that match memories, helping surface relevant information before you make mistakes:

from remembering import recall_hints

# Scan code context for relevant memories
hints = recall_hints("for m in memories: print(m['content'])")

if hints['hints']:
    print("Relevant memories found:")
    for h in hints['hints']:
        print(f"  [{h['type']}] {h['preview']}")
        print(f"    Matched: {h['matched_terms']}")

# Check for unmatched terms (potential new topics)
if hints['unmatched_terms']:
    print(f"New terms: {hints['unmatched_terms']}")

Use explicit terms for targeted lookup:

hints = recall_hints(terms=["muninn", "field", "summary", "content"])
# Returns hints for memories matching any of these terms

Hint structure:

{
    'hints': [
        {
            'memory_id': 'abc-123...',
            'type': 'decision',
            'preview': 'First 100 chars of summary...',
            'matched_terms': ['muninn', 'field'],
            'matched_tags': ['muninn'],
            'priority': 1,
            'relevance_score': 3
        }
    ],
    'term_coverage': {'muninn': ['abc-123'], 'field': ['abc-123', 'def-456']},
    'unmatched_terms': ['content'],
    'warning': None  # or error message if cache unavailable
}

When to use:

• Before writing code that uses recall() - catch field name errors early
• When starting work on a topic - surface forgotten context
• Before making decisions - check for relevant past decisions
• After receiving user instructions - find related memories

Performance: Uses local cache when available (~5ms). Falls back to config-based tag matching.

Edge Cases

Empty recall results: Returns MemoryResultList([]), not an error. Check list length before accessing.

Search literal matching: Current implementation uses SQL LIKE. Searches "API test" matches "API testing" but not "test API" (order matters).

Tag partial matching: tags=["task"] matches memories with tags ["task", "urgent"] via JSON substring search.

Confidence defaults: decision type defaults to 0.8 if not specified. Others default to NULL.

Invalid type: Raises ValueError with list of valid types.

Invalid category: config_set raises ValueError if category not 'profile', 'ops', or 'journal'.

Journal pruning: Call journal_prune() periodically to prevent unbounded growth. Default keeps 40 entries.

Tag mode: tag_mode="all" requires all specified tags to be present. tag_mode="any" (default) matches if any tag present.

Query expansion: When FTS5 returns fewer results than expansion_threshold (default 3), tags are automatically extracted from partial matches and used to find related memories. Set expansion_threshold=0 to disable.

Implementation Notes

• Backend: Turso SQLite HTTP API
• URL: TURSO_URL environment variable or /mnt/project/muninn.env, falls back to default
• Token: TURSO_TOKEN environment variable, /mnt/project/muninn.env, or /mnt/project/turso-token.txt
• Two tables: config (KV) and memories (observations)
• FTS5 search: Porter stemmer tokenizer with BM25 ranking
• HTTP API required (libsql SDK bypasses egress proxy)
• Local SQLite cache for fast recall (< 5ms vs 150ms+ network)
• Thread-safe for background writes