feat(memory): /memory Telegram command (spec 310)

.env.example

@@ -144,6 +144,13 @@ MEMORY_ENABLED=false
 # Models with different dims are rejected at startup. Supporting other
 # dimensions requires changing embedding_model_dims in memory.py.
 #MEMORY_EMBEDDING_MODEL=all-MiniLM-L6-v2
+# Minimum Mem0 similarity score for a memory to be returned. Float in
+# [0.0, 1.0]; default 0.3 matches Mem0's built-in default. Governs both
+# the context-injection path (format_context) and the /memory search UI
+# - one knob, two paths. Raise toward 0.5+ to reduce false positives at
+# the cost of recall; lower toward 0.0 if recall feels missing. Tunable
+# at runtime by editing this file and restarting the service.
+#MEMORY_SEARCH_FLOOR=0.3
 
 # Track 2 Haiku extraction. Requires MEMORY_ENABLED=true and a Claude
 # backend. Off by default at Phase 2 ship (self-reinforcing loop needs

src/kai/bot.py

@@ -54,7 +54,7 @@
     filters,
 )
 
-from kai import github_api, services, sessions, webhook
+from kai import github_api, memory_command, services, sessions, webhook
 from kai.config import (
     DATA_DIR,
     MAX_CONTEXT_CEILING,
@@ -2845,6 +2845,12 @@ async def handle_help(update: Update, context: ContextTypes.DEFAULT_TYPE) -> Non
         "/github add <repo> - Watch a repo\n"
         "/github remove <repo> - Unwatch a repo\n"
         "\n"
+        "/memory - Browse remembered facts by tag\n"
+        "/memory search <q> - Semantic search over memories\n"
+        "/memory stats - Counts and confidence distribution\n"
+        "/memory forget <tag> - Delete all memories with a tag\n"
+        "/memory help - /memory subcommand reference\n"
+        "\n"
         "/voice - Toggle voice on/off\n"
         "/voice only - Voice only (no text)\n"
         "/voice on - Text + voice\n"
@@ -3705,12 +3711,14 @@ def create_bot(config: Config, *, use_webhook: bool = True) -> Application:
     app.add_handler(CommandHandler("voices", handle_voices))
     app.add_handler(CommandHandler("webhooks", handle_webhooks))
     app.add_handler(CommandHandler("github", handle_github))
+    app.add_handler(CommandHandler("memory", memory_command.handle_memory_command))
     app.add_handler(CommandHandler("stop", handle_stop))
 
     # Callback query handlers for inline keyboards (pattern-matched)
     app.add_handler(CallbackQueryHandler(handle_model_callback, pattern=r"^model:"))
     app.add_handler(CallbackQueryHandler(handle_voice_callback, pattern=r"^voice:"))
     app.add_handler(CallbackQueryHandler(handle_workspace_callback, pattern=r"^ws:"))
+    app.add_handler(CallbackQueryHandler(memory_command.handle_memory_callback, pattern=r"^mem:"))
 
     # Media handlers (must be before the catch-all text handler)
     app.add_handler(MessageHandler(filters.PHOTO, handle_photo))

src/kai/config.py

@@ -460,6 +460,19 @@ class Config:
     # an executor thread on a hung subprocess.
     memory_extraction_timeout_s: int = 10
 
+    # Minimum Mem0 similarity score for a memory to be returned by
+    # search-driven paths: both `format_context` (context injection
+    # at session start) and the `/memory search` UI surface in
+    # `memory_command.py`. Values below the floor are dropped before
+    # any ranking. Default 0.3 matches Mem0's built-in default and
+    # the prior hard-coded constant; raise toward 0.5+ to reduce
+    # false positives at the cost of recall. Spec 310 §7.5 documents
+    # the "one knob, two paths" decision: keeping the UI floor and
+    # the context-injection floor in lockstep prevents silent
+    # divergence between "what the user sees in /memory search" and
+    # "what Kai pulls into context" after a config change.
+    memory_search_floor: float = 0.3
+
     def get_workspace_config(self, workspace: Path) -> WorkspaceConfig | None:
         """
         Get per-workspace config for a path, or None for global defaults.
@@ -1400,6 +1413,21 @@ def load_config() -> Config:
     except ValueError:
         raise SystemExit("MEMORY_EXTRACTION_TIMEOUT_S must be an integer") from None
 
+    # Search relevance floor. Float in [0.0, 1.0]; default 0.3 matches
+    # Mem0's built-in default and the prior hard-coded constant. Same
+    # try/except pattern as the other memory_* numeric vars: bad input
+    # exits at startup rather than surfacing as a divide-by-zero or
+    # mysterious "no results" symptom at first query. Range is bounded
+    # at both ends because Mem0 cosine similarity is normalized to
+    # [0.0, 1.0]; a floor outside that range silently filters
+    # everything (>1.0) or nothing (<0.0), both of which are footguns.
+    try:
+        memory_search_floor = float(os.environ.get("MEMORY_SEARCH_FLOOR", "0.3"))
+        if memory_search_floor < 0.0 or memory_search_floor > 1.0:
+            raise SystemExit("MEMORY_SEARCH_FLOOR must be between 0.0 and 1.0")
+    except ValueError:
+        raise SystemExit("MEMORY_SEARCH_FLOOR must be a number") from None
+
     # Per-workspace configuration. Loaded after ALLOWED_WORKSPACES so
     # YAML-defined workspaces can be merged into the allowed set.
     workspace_configs = _load_workspace_configs()
@@ -1557,4 +1585,5 @@ def load_config() -> Config:
         memory_extraction_model=memory_extraction_model,
         memory_extraction_budget_usd=memory_extraction_budget_usd,
         memory_extraction_timeout_s=memory_extraction_timeout_s,
+        memory_search_floor=memory_search_floor,
     )