feat: enable durable memory extraction lifecycle

Author: claude

CHANGELOG.md

@@ -5,6 +5,45 @@ All notable changes to this project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [Unreleased]
+
+## [4.3.0] - 2026-07-04
+
+### Added
+
+- Sessions now resolve a default memory store even when callers do not pass one,
+  preferring an explicit store, then a file memory directory, then configured
+  `memory_dir`, then `<workspace>/.a3s/memory`, with an in-memory fallback and
+  init warning if the file store cannot be created.
+- LLM memory extraction is enabled by default and runs behind a completed-turn
+  significance gate instead of mechanically extracting after every input or tool
+  result. Extraction prompts include related existing memories, including
+  existing `supersedes` / `conflicts_with` relation metadata when available, so
+  the model can consolidate or preserve conflicts. Streaming runs schedule gated
+  extraction in the background after the final response event so UI completion is
+  not blocked by the maintenance LLM call; each memory instance now also runs at
+  most one background extraction at a time to keep maintenance calls bounded.
+
+### Changed
+
+- Successful tool outputs are no longer mechanically written when LLM memory
+  extraction is enabled; failures are still stored immediately.
+- Recalled memory context now carries memory metadata plus concise relation
+  annotations for `supersedes` and `conflicts_with` when present.
+- Memory recall now preserves query-specific search ranking when injecting
+  memories into prompt context, so precise matches are not drowned out by generic
+  high-importance memories. The LLM extraction parser also keeps valid extracted
+  memories when a sibling item is malformed.
+- Session memory now consumes the store's canonical returned item, so store-level
+  duplicate consolidation keeps short-term memory and emitted memory ids aligned
+  with the durable item that actually represents the fact.
+- LLM memory extraction now merges near-duplicate extracted memories into the
+  existing canonical item instead of silently discarding the improved wording,
+  tags, importance, or provenance metadata.
+- Default memory stores now also perform conservative near-duplicate
+  consolidation and conflict-safe pruning, so manually written memories and LLM
+  extracted memories share the same canonical-item lifecycle.
+
 ## [4.2.1] - 2026-06-23
 
 ### Fixed

Cargo.lock

@@ -1,6 +1,6 @@
 [package]
 name = "a3s-code-core"
-version = "4.2.8"
+version = "4.3.0"
 edition = "2021"
 authors = ["A3S Lab Team"]
 license = "MIT"
@@ -14,7 +14,7 @@ path = "src/lib.rs"
 [dependencies]
 # Internal crates
 a3s-common = { version = "0.1.1", path = "../../common" }
-a3s-memory = { version = "0.1.1", path = "../../memory" }
+a3s-memory = { version = "0.1.2", path = "../../memory" }
 a3s-lane = { version = "0.4", path = "../../lane" }
 a3s-search = { version = "1.2.3", path = "../../search", default-features = false, features = ["lightpanda"] }
 

core/Cargo.toml

@@ -39,6 +39,7 @@ mod hook_runtime;
 mod llm_turn;
 mod loop_builder;
 mod loop_runtime;
+mod memory_extraction_runtime;
 mod parallel_tool_runtime;
 mod plan_execution;
 mod planning_runtime;
@@ -106,9 +107,10 @@ pub(crate) struct AgentConfig {
     pub max_parse_retries: u32,
     /// Per-tool execution timeout in milliseconds (`None` = no timeout).
     ///
-    /// When set, each tool execution is wrapped in `tokio::time::timeout`.
-    /// A timeout produces an error result sent back to the LLM rather than
-    /// crashing the session.
+    /// This applies only after permission/HITL approval has completed. HITL
+    /// confirmation waiting is governed by `ConfirmationPolicy` and must not
+    /// consume this tool execution budget. A timeout produces an error result
+    /// sent back to the LLM rather than crashing the session.
     pub tool_timeout_ms: Option<u64>,
     /// Per-model API HTTP timeout in milliseconds (`None` = no timeout).
     ///
@@ -143,7 +145,10 @@ pub(crate) struct AgentConfig {
     /// Maximum context window size in tokens (used for auto-compact calculation).
     /// Default: 200_000.
     pub max_context_tokens: usize,
-    /// Optional agent memory for auto-remember after tool execution and recall before prompts.
+    /// Agent memory for recall and completed-turn extraction.
+    ///
+    /// Session construction resolves a default memory store; this remains
+    /// optional for lower-level/manual `AgentLoop` construction.
     pub memory: Option<Arc<crate::memory::AgentMemory>>,
     /// Inject a continuation message when the LLM stops calling tools before the
     /// task is complete. Enabled by default. Set to `false` to disable.

core/src/agent.rs

@@ -57,6 +57,14 @@ impl AgentLoop {
             .await;
 
         if let Some(sid) = session_id {
+            self.schedule_turn_memory_extraction(
+                state,
+                effective_prompt,
+                &final_text,
+                sid,
+                event_tx,
+            )
+            .await;
             self.notify_turn_complete(sid, effective_prompt, &final_text)
                 .await;
         }