Athena/NexusOS
1
0
Fork 0
forked from enderofwings/NexusOS
Code Pull Requests Activity

feat: sync with upstream — v1.2.0, in-app updates, Projects, modules

Browse Source 
Brings the public tree back in line with the development repo after several
weeks of drift caused by a stale publish include list.

New:
- In-app update path: GET /update/check compares the checkout against
  origin/main and POST /update/apply runs `ncp upgrade` detached (pull,
  rebuild, restart). The sidebar shows the version, checks on click, and
  offers an "update available" pill.
- Projects: a project workspace groups chats and RAG documents, with
  per-project instructions and document retrieval scoped to the active
  project. Replaces the standalone Documents page.
- modules/: auto-discovered feature plugins (mail, network) with their
  frontend counterparts and tests.
- Memory curation runs in-process (synapse/memory/curator.py) on the chat
  model when a conversation goes idle. The separate memory service on :8001
  is gone, along with the launcher lines that started it.

Also: the KDE theme, panel and Promethean terminal assets, the full test
suite, and VERSION 1.2.0.

🤖 Generated with [Claude Code](https://claude.com/claude-code)
This commit is contained in:
janvanwan
2026-08-25 09:13:55 -05:00
parent fe5d18afa7
commit 42eaed647a
88 changed files with 4280 additions and 1888 deletions
Download Patch File Download Diff File Expand all files Collapse all files
synapse/memory/extractor.py
+55 -20
View File
@@ -1,5 +1,12 @@
"""Memory extraction — asks Mistral to evaluate a conversation exchange and
decide if it contains a new permanent personal fact worth saving."""
"""Memory extraction — asks the curator model to read a finished conversation
and decide which new permanent personal facts it contains.
Reads the whole transcript, not a single exchange. A fact is rarely complete in
the turn that introduces it ("I have a Honda" at turn 3 becomes a 2006 Accord
with 312k miles by turn 7), and the store is append-only for the curator, so
extracting per-exchange could only ever accumulate fragments of the same fact.
Waiting until the conversation is idle is also the only reliable signal that a
statement is finished rather than half-said."""
from __future__ import annotations
@@ -17,13 +24,19 @@ _TR = "┅" * 55
_PROMPT = """\
You are a memory curator for a personal AI assistant named Nexus.
Extract EVERY new, permanent personal fact the USER stated in this exchange.
There may be SEVERAL facts in one message — output one JSON object for each.
Extract EVERY new, permanent personal fact the USER stated in this conversation.
There may be SEVERAL facts across the conversation — output one JSON object for
each.
ONLY the USER line is a source of facts. The ASSISTANT line and the existing
memory below are context to help you understand the USER line — never extract
ONLY the USER lines are a source of facts. The ASSISTANT lines and the existing
memory below are context to help you understand the USER lines — never extract
anything from them. If the assistant said it and the user did not, it is NOT a
fact. Copy what the user actually said; do not infer, embellish, or add a
fact.
You are reading the WHOLE conversation, so record the FINAL, most complete form
of each fact. If the user gives a detail early and refines it later, output one
object with the refined version — never one per stage. If the user corrects or
retracts something, keep only what they ended up saying. Copy what the user actually said; do not infer, embellish, or add a
judgement the user did not make (never call something their "favorite",
"main", or "best" unless the user used that word).
@@ -56,8 +69,7 @@ Only invent a new section if none fit, and make it a SHORT single word
(e.g. Hobbies, Pets, Health). Never use a sentence or long phrase as a section.
---
USER: {user_message}
ASSISTANT: {assistant_response}
{transcript}
---
Respond with JSON only — no prose, no markdown fences. Output one object PER
@@ -95,11 +107,11 @@ def _reject_reason(fact: str, user_message: str) -> str | None:
prompt is worded (verified against mistral:7b):
1. Absence claims. It reads the existing-memory block and writes things like
"Jon does not have any pets" - which contradicted four cats already on
"the user does not have any pets" - which contradicted four cats already on
file. Silence is not a fact.
2. Assistant-sourced specifics. It lifts names the ASSISTANT said and
attributes them to the user: a reply that echoed a stale memory row
produced "Jon's main development machine is a MacBook Pro" off the user
produced "the user's main development machine is a MacBook Pro" off the user
message "What am I developing on?".
The grounding test only fires when a fact carries distinctive tokens and
@@ -114,20 +126,44 @@ def _reject_reason(fact: str, user_message: str) -> str | None:
return None
def render_transcript(messages: list[dict]) -> str:
"""The conversation as the curator sees it. Assistant turns are truncated
hard: they are context for reading the user's lines, never a fact source,
and a long reply would otherwise crowd out the lines that matter."""
lines = []
for m in messages:
role = (m.get("role") or "").upper()
if role not in ("USER", "ASSISTANT"):
continue
text = (m.get("content") or "").strip()
if not text:
continue
lines.append(f"{role}: {text[:3000] if role == 'USER' else text[:600]}")
return "\n".join(lines)
async def extract_memory(
user_message: str,
assistant_response: str,
messages: list[dict],
existing_sections: list[str],
existing_texts: list[str],
ollama_manager,
model: str = DEFAULT_MEMORY_MODEL,
num_gpu: int | None = 0,
) -> list[dict]:
"""Ask Mistral to extract saveable memory facts from a conversation exchange.
"""Extract saveable memory facts from a whole conversation.
Returns a list of {"section": ..., "text": ...} — possibly empty. A single
exchange can hold several facts, and Mistral emits one JSON object per fact.
`messages` is the transcript as [{role, content}]. Returns a list of
{"section": ..., "text": ...} — possibly empty. One conversation can hold
several facts, and the model emits one JSON object per fact.
"""
transcript = render_transcript(messages)
if not transcript:
return []
# Grounding is checked against everything the user actually typed, so a fact
# assembled from details spread across several of their turns still passes.
user_text = "\n".join(
(m.get("content") or "") for m in messages if (m.get("role") or "") == "user"
)
if existing_texts:
# ponytail: only the 12 most-recent facts go in the dedup context, not all
# ~40. On a CPU-bound curator (num_gpu=0) prompt-eval dominates, and 40
@@ -148,8 +184,7 @@ async def extract_memory(
prompt = _PROMPT.format(
existing_texts=texts_block,
existing_sections=sections_line,
user_message=user_message[:3000],
assistant_response=assistant_response[:800],
transcript=transcript,
)
# MindTrace: curator pre-flight (full prompt) so its reasoning is visible in
# the same console as the frontline model, not just Python warnings on failure.
@@ -194,7 +229,7 @@ async def extract_memory(
def _keep(o):
if isinstance(o, dict) and o.get("save") and o.get("section") and o.get("text"):
fact = str(o["text"]).strip()
reason = _reject_reason(fact, user_message)
reason = _reject_reason(fact, user_text)
if reason:
_synapse_trace(f"◆ CURATOR DROPPED ({reason}): {fact}\n")
return
@@ -217,7 +252,7 @@ async def extract_memory(
else:
_keep(obj)
if not results:
_log.warning("memory: nothing saved. mistral said: %.300r", text)
_log.warning("memory: nothing saved. curator said: %.300r", text)
_synapse_trace(f"◆ CURATOR VERDICT: nothing to save\n{_TR}\n\n")
else:
_facts = "; ".join(f"[{r['section']}] {r['text']}" for r in results)