Lifecycle Protocols

FAVA Trails exposes lifecycle hooks at key moments — before_save, on_recall, before_propose, and more. Protocols are pre-built hook implementations that plug into these points. Each is independent; enable any combination.

---

SECOM — Extractive Compression

Adapted from SECOM (Tsinghua University and Microsoft, ICLR 2025), which applies LLMLingua-2 extractive compression to conversational memory retrieval. Compresses thoughts at promote time via token-level extraction, keeping the semantic core while dramatically reducing storage.

Hook points: before_propose, before_save, on_recall

Requires the secom extra:

pip install fava-trails[secom]

First use downloads a ~700 MB BERT model from HuggingFace. Pre-download it with:

fava-trails secom warmup

Setup

One command writes the hook entries into your data repo’s config.yaml:

fava-trails secom setup --write

Or add manually:

# config.yaml
lifecycle:
  before_propose:
    - protocol: secom
      config:
        target_ratio: 0.5        # compress to ~50% of original tokens
        force_tokens: []          # tokens to always keep
        model: "microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank"

Key options

Option	Default	Description
target_ratio	0.5	Target compression ratio (0.0–1.0)
force_tokens	[]	Tokens that must survive compression
model	microsoft/llmlingua-2-bert-base-multilingual-cased-meetingbank	HuggingFace model ID

Known limitation

SECOM destroys structured data — JSON, YAML, code blocks, and tables lose their formatting after extraction. Tag thoughts with secom-skip to opt out of compression for structured content.

---

ACE — Agentic Context Engineering

Based on Stanford, UC Berkeley, and SambaNova ACE. Applies playbook-driven reranking and anti-pattern detection to recalled thoughts, surfacing what matters and flagging what’s stale or contradictory.

Hook points: on_startup, on_recall, before_save, after_save, after_propose, after_supersede

Included in the base install — no extras needed.

Setup

fava-trails ace setup --write

Or add manually:

# config.yaml
lifecycle:
  on_recall:
    - protocol: ace
      config:
        scoring: multiplicative    # combine rule scores multiplicatively
        rules_namespace: "preferences/"
        anti_patterns:
          - stale_reference
          - contradictory_belief
          - unsupported_claim

Key options

Option	Default	Description
scoring	multiplicative	How rule scores combine (multiplicative or additive)
rules_namespace	preferences/	Namespace where playbook rules are stored
anti_patterns	[stale_reference, contradictory_belief, unsupported_claim]	Patterns to detect and penalize

How it works

ACE stores scoring rules as thoughts in the preferences/ namespace. When you recall, ACE reranks results by applying each rule multiplicatively — a thought that triggers an anti-pattern gets downweighted, while thoughts matching active playbook rules get boosted. Rules are themselves thoughts, so they participate in supersession and versioning.

---

RLM — MapReduce Orchestration

Based on MIT RLM. Validates mapper outputs, tracks batch progress, and sorts results for reducer consumption — essential for multi-agent workflows that fan out and collect.

Hook points: before_save, after_save, on_recall

Included in the base install — no extras needed.

Setup

fava-trails rlm setup --write

Or add manually:

# config.yaml
lifecycle:
  before_save:
    - protocol: rlm
      config:
        fail_mode: closed          # reject invalid mapper outputs
        required_fields:
          - batch_id
          - mapper_id
        sort_key: "mapper_id"      # sort order for reducer recall

Key options

Option	Default	Description
fail_mode	closed	closed rejects invalid outputs; open warns but allows
required_fields	[batch_id, mapper_id]	Fields every mapper output must contain
sort_key	mapper_id	Field used to sort results for the reducer

How it works

When a mapper agent calls save_thought, RLM validates the metadata contains all required_fields. In closed mode, missing fields cause the save to fail — no partial batches pollute the trail. On recall, RLM sorts results by sort_key so the reducer agent receives them in deterministic order. Batch progress is tracked via metadata, letting orchestrators query completion status.

---

Combining Protocols

Protocols are composable. A common production stack:

lifecycle:
  before_save:
    - protocol: rlm       # validate mapper outputs first
    - protocol: ace        # check anti-patterns before persisting
  before_propose:
    - protocol: secom      # compress before promoting to truth
  on_recall:
    - protocol: ace        # rerank recalled thoughts
    - protocol: rlm        # sort for reducer consumption

Hooks execute in the order listed. If a closed-mode hook rejects a thought, later hooks in the chain are skipped.

---

Further Reading

• FAVA Trails README — full install and MCP setup
• SECOM paper — extractive compression for conversational memory (Tsinghua/Microsoft, ICLR 2025)
• LLMLingua-2 paper — token-level extraction underlying SECOM
• ACE paper — playbook-driven context engineering (Stanford/UC Berkeley/SambaNova, ICLR 2026)
• RLM paper — MapReduce for language model orchestration (MIT)