[feature] summarize_recall — LLM-synthesized recall for context-tight agents

[ramonlimaramos]

Problem

recall returns top-K raw memories. for agents on tight context budgets (haiku, local models, multi-tool chains where memory is one of many tools), 10 raw memories can burn 2-3k tokens. the agent often only needs the synthesis: "what does synapto know about kafka in this project?" — not the verbatim memory texts.

Proposed Solution

new MCP tool summarize_recall that runs the standard hybrid retrieval, then synthesizes the top-K via LLM into a single answer with citations.

signature

summarize_recall(
    query: str,
    tenant: str | None = None,
    max_tokens: int = 500,        # synthesis budget
    k: int = 10,                  # candidates retrieved
    include_citations: bool = True,
) -> SummaryResult

returns:

class SummaryResult:
    summary: str                 # synthesized answer
    citations: list[Citation]    # [memory_id, snippet, score, depth_layer]
    sources_used: int            # how many of the k contributed
    cache_hit: bool

caching

• key = sha256(query | tenant | top_k_memory_ids | model)
• TTL = ephemeral_max_age_hours / 4 (re-synth if memories likely changed)
• stored in redis (already in deps)
• bypass with force_refresh=True

LLM provider

• reuse synapto.embeddings.registry pattern: pluggable provider
• default to claude-haiku-4-5 (fast + cheap)
• fall back to disabled-with-error if no API key (this tool requires LLM)

prompt template

system prompt enforces: "only synthesize from provided memories, cite memory_id inline, refuse if memories don't support an answer". reduces hallucination risk.

Trade-offs

• cost: each call = LLM invocation. mitigation: aggressive cache, opt-in tool (agent picks recall vs summarize_recall).
• fidelity loss: summarization can lose nuance. mitigation: always return citations so caller can drill into raw if needed.
• latency: adds ~500-1500ms vs raw recall. acceptable for context-saving usecase.
• dependency: pulls anthropic SDK as optional extra (synapto[summarize]).

Out of scope

• streaming summaries (return-and-be-done is fine for MCP)
• cross-tenant summarization (same tenant only)

Success criteria

• agents on haiku see meaningful token savings on memory-heavy chains
• citations are accurate (every claim in summary maps to a memory_id)
• cache hit rate > 60% on stable topics in a session

References

• byterover ships 6 compression strategies for context optimization — confirms agents want this