Memory Layers

Sources: Antigravity, 2026-07-05 Raw: Memory Layers Source

Memory Layers

Four-layer memory system with automatic PII/PHI redaction before anything persists.

Layer	Module	Scope	Persistence
1. Working Memory	`context.py`	Per LLM call	Never
2. Session State	ADK `session.state`	Per conversation	With DB-backed service
3. Long-Term Memory	`memory.py` → MemoryService	Cross-session	With Vertex AI Memory Bank
4. A2A Context	`prepare_a2a_context` → RemoteA2aAgent	Per delegation	Never

Layer 1 — Working Memory (context engineering)

context.py assembles the per-call context window. It never calls the LLM — compaction is rule-based (truncation + metadata preservation).

Function	Purpose
`estimate_tokens(text)`	chars/4 heuristic for budget decisions (no tokenizer)
`build_structured_context(role, environment, task, constraints)`	XML-delimited sections so the model separates instructions from data
`compact_history(turns, keep_recent=10, max_total_tokens=8000)`	Summarizes old turns to metadata, preserves recent verbatim; append-only (returns new list)
`inject_at_boundaries(context, critical_info)`	Places critical info at start AND end ("lost in the middle" mitigation)

Token pipeline: Collect → Rank → Compress → Budget → Assemble. config.deterministic_json() (sorted keys, compact separators) improves KV-cache hits.

Layer 2 — Session State

output_key plumbing carries data between pipeline stages — see [[Agent Architecture]].
State prefixes: no prefix (session), user: (cross-session), app: (global), temp: (invocation-only, never persisted). Layer 2 rate limiting uses temp: state ([[Security Layers]]).

Layer 3 — Long-Term Memory

InMemoryMemoryService (dev) / VertexAiMemoryBankService (prod), selected by environment in memory.py.
Auto-saved via after_agent_callback when sessions complete.
PHI filtered before storage (detect_phi + redact_phi).
Written by pipeline audit stages: save_qa_to_memory, save_query_to_memory; recalled by the orchestrator via search_past_conversations.

Layer 4 — A2A Context

Only task-relevant, non-sensitive data crosses agent boundaries: no PII/PHI, no secrets, no temp:/user: keys. Used by RemoteA2aAgent delegations — see [[MCP and A2A]].

Related: [[Security Layers]] · [[Testing and Eval]] (test_memory.py, test_context.py)