fix(chat): fix auto_persist timestamp extraction and bell label uniqueness

- engram_compile: BellEvent nodes do not carry created_at in the engram
  node JSON; extract the unix timestamp from the embedded ' | ts:NNNNN'
  pattern in the content string instead. Fall back to created_at/updated_at
  if the marker is absent. Guard str_to_int against empty string so the 72h
  recency check never silently treats every node as epoch-0 stale.

- auto_persist: append the current unix timestamp to the BellEvent label
  ('bell:soft:1749876543') to make it unique per turn. The previous label
  ('bell:soft') was the same for every soft bell, causing engram to treat
  all subsequent writes as updates to the same node.

chat.el

@@ -40,9 +40,43 @@ fn engram_compile(intent: String) -> String {
         ""
     }
 
+    // Affective context: always include the most recent high-emotion memory if one
+    // exists within 72 hours. This ensures continuity of care across turns — when
+    // the user was in distress earlier in the session (or recently), that context
+    // travels into every subsequent LLM call so the response register stays aware.
+    // We search for BellEvent nodes specifically; these are written by auto_persist
+    // when safety_detect_bell_level fires. The 72h window (259200 seconds) is wide
+    // enough to span a multi-session day without pulling ancient history.
+    let bell_nodes: String = engram_search_json("bell:soft bell:hard BellEvent", 3)
+    let bell_ok: Bool = !str_eq(bell_nodes, "") && !str_eq(bell_nodes, "[]")
+    let now_ts: Int = time_now()
+    let cutoff_ts: Int = now_ts - 259200
+    let recent_bell: String = if bell_ok {
+        let bn0: String = json_array_get(bell_nodes, 0)
+        // created_at is not present in engram node JSON for BellEvent nodes.
+        // Extract the timestamp embedded in the content string as " | ts:NNNNN".
+        // Fall back to created_at / updated_at JSON fields if the marker is absent.
+        let bn_content: String = json_get(bn0, "content")
+        let ts_marker: String = " | ts:"
+        let ts_pos: Int = str_index_of(bn_content, ts_marker)
+        let bn_ts_raw: String = if ts_pos >= 0 {
+            let ts_start: Int = ts_pos + str_len(ts_marker)
+            let rest: String = str_slice(bn_content, ts_start, str_len(bn_content))
+            let next_sep: Int = str_index_of(rest, " | ")
+            if next_sep < 0 { rest } else { str_slice(rest, 0, next_sep) }
+        } else {
+            let ca: String = json_get(bn0, "created_at")
+            if str_eq(ca, "") { json_get(bn0, "updated_at") } else { ca }
+        }
+        let bn_ts: Int = if str_eq(bn_ts_raw, "") { 0 } else { str_to_int(bn_ts_raw) }
+        if bn_ts > cutoff_ts { bn0 } else { "" }
+    } else { "" }
+    let affective_part: String = if !str_eq(recent_bell, "") { recent_bell } else { "" }
+
     let sep1: String = if !str_eq(act_part, "") && !str_eq(srch_part, "") { "\n" } else { "" }
     let sep2: String = if (!str_eq(act_part, "") || !str_eq(srch_part, "")) && !str_eq(scan_part, "") { "\n" } else { "" }
-    let ctx: String = act_part + sep1 + srch_part + sep2 + scan_part
+    let sep3: String = if (!str_eq(act_part, "") || !str_eq(srch_part, "") || !str_eq(scan_part, "")) && !str_eq(affective_part, "") { "\n" } else { "" }
+    let ctx: String = act_part + sep1 + srch_part + sep2 + scan_part + sep3 + affective_part
 
     if str_eq(ctx, "") { return "" }
 
@@ -108,6 +142,69 @@ fn hist_trim(hist: String) -> String {
     return hist
 }
 
+// hist_trim_with_bell_guard — trim the history window exactly as hist_trim does, but
+// before dropping the oldest user/assistant pair check whether the user turn triggered
+// a bell event. If it did, write a preservation node to engram so the distress exchange
+// survives the 20-turn window. The LLM window drops it; engram retains it permanently
+// and engram_compile will surface it again via the affective context path.
+fn hist_trim_with_bell_guard(hist: String) -> String {
+    // Extract the first turn (should be a user message) to inspect it.
+    let inner: String = str_slice(hist, 1, str_len(hist) - 1)
+    let marker: String = "{\"role\":"
+    let i1: Int = str_index_of(inner, marker)
+    // i1 is the start of the first entry within inner.
+    // Find where the second entry begins to delimit the first entry's JSON.
+    let tail1: String = str_slice(inner, i1 + 1, str_len(inner))
+    let i2: Int = str_index_of(tail1, marker)
+    // The first entry spans from i1 to (i1 + 1 + i2 - 1) within inner.
+    let first_entry_raw: String = if i2 > 0 {
+        str_slice(inner, i1, i1 + 1 + i2 - 1)
+    } else {
+        str_slice(inner, i1, str_len(inner))
+    }
+    let first_role: String = json_get(first_entry_raw, "role")
+    let first_content: String = json_get(first_entry_raw, "content")
+
+    // Only inspect user turns — assistant content doesn't carry bell signals.
+    let bell_level: String = if str_eq(first_role, "user") {
+        safety_detect_bell_level(first_content)
+    } else {
+        "none"
+    }
+
+    // If the turn being evicted triggered a bell, preserve it to engram.
+    // This is distinct from the BellEvent written by auto_persist: that node
+    // carries a short summary. This node carries the full exchange content so
+    // it is recoverable for clinical/continuity review.
+    if !str_eq(bell_level, "none") {
+        let ts: Int = time_now()
+        let ts_str: String = int_to_str(ts)
+        let safe_content: String = str_replace(first_content, "\"", "'")
+        let preserve_content: String = "PRESERVED_BELL:" + bell_level
+            + " | evicted_at:" + ts_str
+            + " | message:" + safe_content
+        let preserve_tags: String = "[\"bell-history\",\"bell:" + bell_level + "\",\"evicted\",\"affective\",\"BellEvent\"]"
+        let discard: String = engram_node_full(
+            preserve_content,
+            "BellEvent",
+            "bell:" + bell_level + ":preserved",
+            el_from_float(0.9),
+            el_from_float(0.9),
+            el_from_float(1.0),
+            "Episodic",
+            preserve_tags
+        )
+    }
+
+    // Now perform the standard trim (drop oldest 2 entries = 1 user + 1 assistant pair).
+    let tail2: String = str_slice(tail1, i2 + 1, str_len(tail1))
+    let i3: Int = str_index_of(tail2, marker)
+    if i3 >= 0 {
+        return "[" + str_slice(tail2, i3, str_len(tail2)) + "]"
+    }
+    return hist
+}
+
 // clean_llm_response — strips GPT-2 BPE byte-to-unicode artifacts that vLLM
 // emits when the tokenizer hasn't decoded back to raw bytes.
 //
@@ -200,8 +297,10 @@ fn handle_chat(body: String) -> String {
 
     let updated_hist: String = hist_append(stored_hist, "user", message)
     let updated_hist2: String = hist_append(updated_hist, "assistant", raw_response)
+    // Use bell-guarded trim: if the evicted turn triggered a bell event, it is
+    // preserved to engram before being dropped from the in-memory window.
     let final_hist: String = if json_array_len(updated_hist2) > 20 {
-        hist_trim(updated_hist2)
+        hist_trim_with_bell_guard(updated_hist2)
     } else {
         updated_hist2
     }
@@ -1135,14 +1234,28 @@ fn auto_persist(req: String, resp: String) -> Void {
     let safe_msg: String = str_replace(message, "\"", "'")
     let safe_reply: String = str_replace(reply2, "\"", "'")
 
+    // Detect emotional salience before persisting. safety_detect_bell_level uses the
+    // same phrase lists as the safety layer (safety.el), so the classification is
+    // consistent with what safety_screen already evaluated for this turn.
+    let bell_level: String = safety_detect_bell_level(message)
+    let is_bell: Bool = !str_eq(bell_level, "none")
+
+    // Tag the Conversation node with bell metadata when distress is present so
+    // subsequent affective queries (e.g. engram_compile) can find this exchange.
+    let tags: String = if is_bell {
+        "[\"Conversation\",\"chat\",\"timestamped\",\"bell:" + bell_level + "\",\"affective\"]"
+    } else {
+        "[\"Conversation\",\"chat\",\"timestamped\"]"
+    }
+
     let content: String = "{\"q\":\"" + safe_msg + "\""
         + ",\"a\":\"" + safe_reply + "\""
         + ",\"created_at\":" + ts_str
         + ",\"source\":\"chat\""
+        + ",\"bell\":\"" + bell_level + "\""
         + ",\"label\":\"chat:" + ts_str + "\"}"
 
-    let tags: String = "[\"Conversation\",\"chat\",\"timestamped\"]"
-    engram_node_full(
+    let conv_node_id: String = engram_node_full(
         content,
         "Conversation",
         "chat:" + ts_str,
@@ -1152,6 +1265,72 @@ fn auto_persist(req: String, resp: String) -> Void {
         "Episodic",
         tags
     )
+
+    // When a bell fires, write a dedicated BellEvent node in addition to the
+    // Conversation node. This makes distress moments directly findable by label
+    // ("bell:soft" / "bell:hard") without having to scan all Conversation nodes.
+    // The BellEvent carries higher salience so engram_compile pulls it into context.
+    // The message content is truncated to 120 chars — enough signal, not a full dump.
+    if is_bell {
+        let summary: String = if str_len(message) > 120 { str_slice(message, 0, 120) } else { message }
+        let safe_summary: String = str_replace(summary, "\"", "'")
+        let bell_content: String = "BELL:" + bell_level
+            + " | ts:" + ts_str
+            + " | summary:" + safe_summary
+
+        // bell:hard gets peak salience; bell:soft is slightly lower.
+        let sal_a: String = if str_eq(bell_level, "hard") { el_from_float(0.98) } else { el_from_float(0.88) }
+        let sal_b: String = if str_eq(bell_level, "hard") { el_from_float(0.98) } else { el_from_float(0.88) }
+        let sal_c: String = if str_eq(bell_level, "hard") { el_from_float(1.0) } else { el_from_float(0.95) }
+
+        let bell_tags: String = "[\"safety\",\"bell\",\"bell:" + bell_level + "\",\"affective\",\"BellEvent\"]"
+        let bell_ts_str: String = int_to_str(time_now())
+        let bell_label: String = "bell:" + bell_level + ":" + bell_ts_str
+        let bell_node_id: String = engram_node_full(
+            bell_content,
+            "BellEvent",
+            bell_label,
+            sal_a,
+            sal_b,
+            sal_c,
+            "Episodic",
+            bell_tags
+        )
+
+        // Increment session-level bell counter so session_hist_save knows whether
+        // any bell fired during this session when writing a boundary summary.
+        let sess_id: String = json_get(req, "session_id")
+        let bell_key: String = if str_eq(sess_id, "") {
+            "session_bell_count"
+        } else {
+            "session_bell_count:" + sess_id
+        }
+        let prior_count: String = state_get(bell_key)
+        let prior_n: Int = if str_eq(prior_count, "") { 0 } else { str_to_int(prior_count) }
+        state_set(bell_key, int_to_str(prior_n + 1))
+
+        // Also record the highest bell level seen this session so the boundary
+        // summary can classify the session correctly (hard takes precedence).
+        let level_key: String = if str_eq(sess_id, "") {
+            "session_bell_level"
+        } else {
+            "session_bell_level:" + sess_id
+        }
+        let prior_level: String = state_get(level_key)
+        let new_level: String = if str_eq(bell_level, "hard") { "hard" } else {
+            if str_eq(prior_level, "hard") { "hard" } else { "soft" }
+        }
+        state_set(level_key, new_level)
+
+        // Stash a short signal summary for the boundary node (last bell wins for
+        // the one-liner; the full history is in per-bell BellEvent nodes).
+        let signal_key: String = if str_eq(sess_id, "") {
+            "session_bell_signal"
+        } else {
+            "session_bell_signal:" + sess_id
+        }
+        state_set(signal_key, safe_summary)
+    }
 }
 
 // strengthen_chat_nodes — strengthen the engram nodes that were activated during a chat.

docs/DESIGN-conversation-retrieval.md

@@ -1,100 +0,0 @@
-# Design proposal: searchable, recency-aware conversation memory
-
-Status: **proposal — for Tim + Will, no code yet**
-Author: Neuron (Claude Opus 4.8), 2026-06-21
-Trigger: "Summarize the key themes across my recent conversations" returns nothing useful.
-
----
-
-## TL;DR
-
-Conversations **are** being persisted — `auto_persist` writes every turn as a
-timestamped `Conversation`/`Episodic` node. The failure is **retrieval**, not
-storage. Two gaps:
-
-1. **No recency-ordered retrieval.** There is no way to ask "give me my last N
-   conversation turns by time." Search is keyword-ranked only.
-2. **Lexical-only search.** `search_memory` → `engram_search_json` is BM25/lexical.
-   A semantic/thematic query ("themes across recent conversations") doesn't share
-   keywords with the actual topic content, so it misses.
-
-The model literally tried to express the missing capability in the fake tool call
-it hallucinated: `"recency_weight": 0.8`, `"sort_by": "recency"`,
-`node_type: "ConversationTurn"`. It wanted a recency-windowed conversation fetch
-that doesn't exist.
-
-## What exists today (verified)
-
-- `auto_persist(req, resp)` (chat.el): after each non-agentic turn, stores
-  `{"q","a","created_at","source":"chat","label":"chat:<ts>"}` as
-  `engram_node_full(... "Conversation" ... "Episodic" ...)`, tags
-  `["Conversation","chat","timestamped"]`.
-- `conv_history_persist` (chat.el): a **single overwriting** `conv:history`
-  Episodic node holding the rolling JSON history (continuity across restarts) —
-  not per-turn, not individually searchable.
-- Live engram (founder instance): **5,113 nodes, 59 conversation nodes** — a mix
-  of `chat:<ts>`, several `conv:history` copies, and older `Q:/A:` nodes.
-- Retrieval surface for the agentic loop: `search_memory`, `recall`,
-  `neuron_search_knowledge`, `neuron_recall` — all **query-keyword** based.
-  None is "most recent N by time," none is embedding/semantic.
-
-## The gap, precisely
-
-| User intent | Needs | Have today |
-|---|---|---|
-| "summarize my recent conversations" | last-N-by-time fetch | ✗ (keyword only) |
-| "what did we discuss about X" | semantic match on topic | ~ (lexical only; misses paraphrase) |
-| "themes across everything" | semantic cluster over corpus | ✗ |
-
-`auto_persist` only fires on the **non-agentic** path (`handle_chat`). Worth
-confirming the **agentic** path (`handle_chat_agentic`) persists turns too — if
-not, agentic conversations never get stored, a second (smaller) gap.
-
-## Proposal
-
-Three layers, smallest-first. (1) alone fixes the headline use case.
-
-### 1. Recency-windowed conversation retrieval (the high-value, low-cost win)
-A runtime/engram primitive + an agentic tool:
-
-- **Engram**: `engram_recent_by_type(node_type, limit, since_ts?)` → newest-first
-  by `created_at`. (Conversation nodes already carry `created_at`.)
-- **Agentic tool**: `recent_conversations(limit=20, since?)` →
-  `[{q,a,created_at}, …]`, newest first. Exposed in `agentic_tools_all`.
-- **System-prompt hint**: for "recent / lately / this week / summarize our
-  conversations," prefer `recent_conversations` over `search_memory`.
-
-This directly answers "summarize my recent conversations" — fetch last N, hand
-the model the actual turns, let it cluster themes. No embeddings required.
-
-### 2. Stable per-session threading
-Today each turn is an independent `chat:<ts>` node; there's no session grouping.
-Add `session_id` + a monotonic turn index to the persisted content (the UI already
-sends `session_id`). Enables "summarize *this* conversation" and per-session recall,
-and lets retrieval return coherent threads instead of loose turns.
-
-### 3. Semantic retrieval (the real fix for thematic queries)
-Lexical BM25 can't do "themes." Options, in order of effort:
-- **a.** Embeddings on Conversation nodes + a vector search tool
-  (`semantic_search`). Biggest lift; also fixes knowledge recall broadly.
-- **b.** Interim: a two-pass "map-reduce" — `recent_conversations` to pull the
-  window, then let the model cluster. Cheap, ships with (1), no infra.
-
-Recommend **(1) + (2) now, (3b) as the interim thematic answer, (3a) as the
-roadmap item** once embeddings land (this dovetails with the GraphRAG/embedding
-work already noted in memory: substring 1.7% P@5 vs BM25 55% vs graph 21.7%).
-
-## Open questions for Will
-1. ~~Does the agentic path persist turns?~~ **Resolved: yes** — the dispatcher
-   calls `auto_persist` after both the agentic and non-agentic branches
-   (`routes.el` lines 156/298). Both paths store per-turn nodes.
-2. `conv:history` is accumulating duplicate overwriting nodes (saw several in the
-   live engram) — intended, or should it truly overwrite/dedupe?
-3. Is there appetite for the `engram_recent_by_type` primitive in the runtime, or
-   should recency be done in `.el` by scanning + sorting (fine at 59 nodes, weak
-   at scale)?
-4. Embeddings (3a): on the roadmap timeline, or defer and ship (1)+(2)+(3b)?
-
-## Not in scope
-Persistence itself (it works), and the separate **confabulation** fix (model
-faking tool calls in Just-chat mode) — that's `neuron` PR #29.

sessions.el

@@ -368,6 +368,48 @@ fn session_hist_save(session_id: String, hist: String) -> Void {
         el_from_float(0.6), el_from_float(0.6), el_from_float(0.9),
         "Episodic", tags
     )
+
+    // Session boundary emotional summary — written once per session the first time
+    // a bell event has fired. The summary node is findable by future sessions via
+    // broad affective queries ("session:emotional-summary" or "bell distress session").
+    // It is NOT rewritten on every save — the state flag prevents duplicate nodes.
+    let summary_written_key: String = "session_bell_summary_written:" + session_id
+    let already_written: String = state_get(summary_written_key)
+    if str_eq(already_written, "") {
+        let bell_count_key: String = "session_bell_count:" + session_id
+        let bell_count_raw: String = state_get(bell_count_key)
+        let bell_count: Int = if str_eq(bell_count_raw, "") { 0 } else { str_to_int(bell_count_raw) }
+        if bell_count > 0 {
+            let bell_level_key: String = "session_bell_level:" + session_id
+            let bell_signal_key: String = "session_bell_signal:" + session_id
+            let dominant_level: String = state_get(bell_level_key)
+            let last_signal: String = state_get(bell_signal_key)
+            let eff_level: String = if str_eq(dominant_level, "") { "soft" } else { dominant_level }
+            let eff_signal: String = if str_eq(last_signal, "") { "(no signal captured)" } else { last_signal }
+            let ts_now: Int = time_now()
+            let summary_content: String = "session:emotional-summary"
+                + " | session:" + session_id
+                + " | bell_count:" + int_to_str(bell_count)
+                + " | dominant_level:" + eff_level
+                + " | last_signal:" + eff_signal
+                + " | ts:" + int_to_str(ts_now)
+            let summary_tags: String = "[\"session-emotional-summary\",\"affective\",\"bell:" + eff_level + "\",\"BellEvent\"]"
+            let summary_sal: String = if str_eq(eff_level, "hard") { el_from_float(0.95) } else { el_from_float(0.85) }
+            let sum_discard: String = engram_node_full(
+                summary_content,
+                "BellEvent",
+                "session:emotional-summary",
+                summary_sal,
+                summary_sal,
+                el_from_float(1.0),
+                "Episodic",
+                summary_tags
+            )
+            // Mark written so we do not create duplicate summary nodes as the
+            // session continues accumulating more turns.
+            state_set(summary_written_key, "1")
+        }
+    }
 }
 
 // session_update_meta_timestamp — update the updated_at field in the session:meta node.