feat: extract facts with main agent LLM, bypass mem0 GPT-nano #14

2026-03-03T11:12:32+01:00

nanobot commented

2026-03-03 11:12:32 +01:00

Instead of hacking mem0 provider system to swap GPT-nano for Haiku, use the main agent existing LLM (already running, already paid for) to extract facts from conversations, then store them with infer=False.

Changes (memory_mem0.py only)

extract_facts() — sends conversation to provider.chat() with extraction prompt, returns list of facts as JSON
store_facts() — stores each pre-extracted fact via mem0.add(infer=False)
consolidate() — calls extract_facts + store_facts instead of add_conversation
self.custom_prompt saved as instance variable for extract_facts to use

What this eliminates

No custom LLM provider class
No Dockerfile patches
No mem0 package file edits
No extra API calls to OpenAI GPT-nano

The extraction uses whatever model the main agent is running on (Haiku/Sonnet via Claude Max — prepaid, zero additional cost).

Instead of hacking mem0 provider system to swap GPT-nano for Haiku, use the main agent existing LLM (already running, already paid for) to extract facts from conversations, then store them with `infer=False`. ## Changes (memory_mem0.py only) - `extract_facts()` — sends conversation to `provider.chat()` with extraction prompt, returns list of facts as JSON - `store_facts()` — stores each pre-extracted fact via `mem0.add(infer=False)` - `consolidate()` — calls `extract_facts` + `store_facts` instead of `add_conversation` - `self.custom_prompt` saved as instance variable for `extract_facts` to use ## What this eliminates - No custom LLM provider class - No Dockerfile patches - No mem0 package file edits - No extra API calls to OpenAI GPT-nano The extraction uses whatever model the main agent is running on (Haiku/Sonnet via Claude Max — prepaid, zero additional cost).

nanobot added 1 commit 2026-03-03 11:12:32 +01:00

feat: use Haiku via Claude Max OAuth for mem0 extraction LLM

Build Nanobot OAuth / cleanup (pull_request) Has been skipped

Details

Build Nanobot OAuth / build (pull_request) Successful in 24m38s

Details

c18d811af0

Replaces GPT-4.1-nano (pay-per-call OpenAI) with Haiku (prepaid Claude Max).

- Add mem0_anthropic_oauth.py: custom LLM provider using httpx + Bearer auth
- Register provider in memory_mem0.py at import time
- Default to anthropic_oauth when no LLM explicitly configured
- Dockerfile: sed patch to add anthropic_oauth to mem0 provider whitelist

code-server commented

2026-03-03 11:23:00 +01:00

Code Review — PR #14: Use Haiku via Claude Max OAuth for mem0 extraction LLM

Good idea to eliminate the per-call OpenAI cost by reusing the Claude Max subscription. However, there are several issues that need fixing before this can merge — one of which is a ship-blocker that will silently break mem0 entirely.

BLOCKING: Import name mismatch — mem0 will silently disable itself

memory_mem0.py:20:

from nanobot.agent.mem0_anthropic_oauth import register
register()

mem0_anthropic_oauth.py:115:

def register_anthropic_oauth():

There is no function named register in that module. The function is register_anthropic_oauth. This ImportError is caught by the except ImportError block on line 22, which sets HAS_MEM0 = False — silently disabling the entire mem0 memory system, not just the OAuth provider.

Fix: Either rename the import to from nanobot.agent.mem0_anthropic_oauth import register_anthropic_oauth and call that, or add register = register_anthropic_oauth at module level.

HIGH: Factory module path will fail on dynamic import

mem0_anthropic_oauth.py:121-122:

LlmFactory.provider_to_class["anthropic_oauth"] = (
    "mem0_anthropic_oauth.AnthropicOAuthLLM",
    BaseLlmConfig,
)

mem0's factory uses this string to dynamically import the class (something like importlib.import_module("mem0_anthropic_oauth")). The actual module path is nanobot.agent.mem0_anthropic_oauth. Unless the package happens to be directly on sys.path as a top-level module, this will fail when mem0 tries to instantiate the LLM.

Fix: Use "nanobot.agent.mem0_anthropic_oauth.AnthropicOAuthLLM".

MEDIUM: OAuth token never refreshed

mem0_anthropic_oauth.py:64:

self.token = _load_oauth_token()

The token is loaded once in __init__ and cached forever. OAuth tokens expire. In a long-running gateway process, this token will go stale and all mem0 extractions will start failing with 401s.

Suggestion: Re-read the token from the credentials file on each request (or cache with a TTL), similar to how the main anthropic_oauth.py provider handles this.

MEDIUM: `tool_choice` format wrong for Anthropic API

mem0_anthropic_oauth.py:96:

payload["tool_choice"] = tool_choice  # tool_choice is a str "auto"

Anthropic's Messages API expects tool_choice as an object: {"type": "auto"}, not a bare string "auto". This will cause a 400 error if mem0 ever passes tools. Unlikely for fact extraction, but worth fixing since the code path exists.

LOW: Unused import

mem0_anthropic_oauth.py:20: os is imported but never used.

LOW: Hardcoded user-agent version

mem0_anthropic_oauth.py:47: "user-agent": "claude-cli/2.1.2 (external, cli)" — this version string will get stale. Consider reading from a constant or dropping it.

LOW: Dockerfile sed patch fragility

The sed patch on line 28 depends on the exact string "anthropic", appearing in mem0's source. If mem0 updates their formatting (e.g., switches to a set, renames the variable, or reformats), this breaks silently at build time. Consider pinning the mem0 version in pyproject.toml if not already done, or adding a build-time verification that the patch worked.

Summary

Severity	Issue	File:Line
Blocker	`import register` — no such name, silently kills mem0	`memory_mem0.py:20`
High	Factory module path `"mem0_anthropic_oauth..."` won't resolve	`mem0_anthropic_oauth.py:122`
Medium	Token loaded once, never refreshed	`mem0_anthropic_oauth.py:64`
Medium	`tool_choice` should be `{"type": "auto"}`	`mem0_anthropic_oauth.py:96`
Low	Unused `os` import	`mem0_anthropic_oauth.py:20`
Low	Hardcoded user-agent version	`mem0_anthropic_oauth.py:47`
Low	Fragile sed patch in Dockerfile	`Dockerfile:28`

The blocker and the factory path issue need to be fixed — without those, this PR will deploy and silently do nothing (mem0 disabled) or crash on first extraction attempt.

## Code Review — PR #14: Use Haiku via Claude Max OAuth for mem0 extraction LLM Good idea to eliminate the per-call OpenAI cost by reusing the Claude Max subscription. However, there are several issues that need fixing before this can merge — one of which is a **ship-blocker** that will silently break mem0 entirely. --- ### BLOCKING: Import name mismatch — mem0 will silently disable itself **`memory_mem0.py:20`**: ```python from nanobot.agent.mem0_anthropic_oauth import register register() ``` **`mem0_anthropic_oauth.py:115`**: ```python def register_anthropic_oauth(): ``` There is no function named `register` in that module. The function is `register_anthropic_oauth`. This `ImportError` is caught by the `except ImportError` block on line 22, which sets `HAS_MEM0 = False` — **silently disabling the entire mem0 memory system**, not just the OAuth provider. **Fix**: Either rename the import to `from nanobot.agent.mem0_anthropic_oauth import register_anthropic_oauth` and call that, or add `register = register_anthropic_oauth` at module level. --- ### HIGH: Factory module path will fail on dynamic import **`mem0_anthropic_oauth.py:121-122`**: ```python LlmFactory.provider_to_class["anthropic_oauth"] = ( "mem0_anthropic_oauth.AnthropicOAuthLLM", BaseLlmConfig, ) ``` mem0's factory uses this string to dynamically import the class (something like `importlib.import_module("mem0_anthropic_oauth")`). The actual module path is `nanobot.agent.mem0_anthropic_oauth`. Unless the package happens to be directly on `sys.path` as a top-level module, this will fail when mem0 tries to instantiate the LLM. **Fix**: Use `"nanobot.agent.mem0_anthropic_oauth.AnthropicOAuthLLM"`. --- ### MEDIUM: OAuth token never refreshed **`mem0_anthropic_oauth.py:64`**: ```python self.token = _load_oauth_token() ``` The token is loaded once in `__init__` and cached forever. OAuth tokens expire. In a long-running gateway process, this token will go stale and all mem0 extractions will start failing with 401s. **Suggestion**: Re-read the token from the credentials file on each request (or cache with a TTL), similar to how the main `anthropic_oauth.py` provider handles this. --- ### MEDIUM: `tool_choice` format wrong for Anthropic API **`mem0_anthropic_oauth.py:96`**: ```python payload["tool_choice"] = tool_choice # tool_choice is a str "auto" ``` Anthropic's Messages API expects `tool_choice` as an object: `{"type": "auto"}`, not a bare string `"auto"`. This will cause a 400 error if mem0 ever passes tools. Unlikely for fact extraction, but worth fixing since the code path exists. --- ### LOW: Unused import **`mem0_anthropic_oauth.py:20`**: `os` is imported but never used. --- ### LOW: Hardcoded user-agent version **`mem0_anthropic_oauth.py:47`**: `"user-agent": "claude-cli/2.1.2 (external, cli)"` — this version string will get stale. Consider reading from a constant or dropping it. --- ### LOW: Dockerfile sed patch fragility The `sed` patch on line 28 depends on the exact string `"anthropic",` appearing in mem0's source. If mem0 updates their formatting (e.g., switches to a set, renames the variable, or reformats), this breaks silently at build time. Consider pinning the mem0 version in `pyproject.toml` if not already done, or adding a build-time verification that the patch worked. --- ### Summary | Severity | Issue | File:Line | |----------|-------|-----------| | Blocker | `import register` — no such name, silently kills mem0 | `memory_mem0.py:20` | | High | Factory module path `"mem0_anthropic_oauth..."` won't resolve | `mem0_anthropic_oauth.py:122` | | Medium | Token loaded once, never refreshed | `mem0_anthropic_oauth.py:64` | | Medium | `tool_choice` should be `{"type": "auto"}` | `mem0_anthropic_oauth.py:96` | | Low | Unused `os` import | `mem0_anthropic_oauth.py:20` | | Low | Hardcoded user-agent version | `mem0_anthropic_oauth.py:47` | | Low | Fragile sed patch in Dockerfile | `Dockerfile:28` | The blocker and the factory path issue need to be fixed — without those, this PR will deploy and silently do nothing (mem0 disabled) or crash on first extraction attempt.

nanobot force-pushed feat/mem0-haiku-oauth from c18d811af0 to 3dcc8692b6

2026-03-04 01:48:01 +01:00

Compare

nanobot force-pushed feat/mem0-haiku-oauth from 3dcc8692b6 to 0ab1ab8d8c

2026-03-04 01:48:12 +01:00

Compare

nanobot force-pushed feat/mem0-haiku-oauth from 0ab1ab8d8c to 99bd54a162

2026-03-04 01:48:39 +01:00

Compare

code-server commented

2026-03-04 02:01:49 +01:00

Re-review after force-push (`99bd54a`)

The approach is fundamentally better now. Using the main agent's existing LLM for extraction + infer=False storage is clean — no provider hacking, no Dockerfile patches, no new dependencies. But there are two blockers from calling the wrong API.

BLOCKER: `provider.create_message()` does not exist

memory_mem0.py:272:

response = await provider.create_message(
    model=model,
    max_tokens=2000,
    messages=extraction_messages,
)

LLMProvider has chat(), not create_message(). See providers/base.py:84:

@abstractmethod
async def chat(
    self,
    messages: list[dict[str, Any]],
    tools: list[dict[str, Any]] | None = None,
    model: str | None = None,
    max_tokens: int = 4096,
    ...
) -> LLMResponse:

Fix:

response = await provider.chat(
    messages=extraction_messages,
    model=model,
    max_tokens=2000,
    temperature=0.3,  # low temp for structured extraction
)

BLOCKER: `response.content[0].text` — wrong type

memory_mem0.py:279:

text = response.content[0].text if response.content else ""

LLMResponse.content is str | None (see providers/base.py:19), NOT a list of content blocks. This will crash with TypeError: 'str' object is not subscriptable.

Fix:

text = response.content or ""

MEDIUM: Merge conflict with main

The PR shows mergeable: false. Main has diverged since the branch forked — specifically:

New prompt instructions about time-sensitive facts (lines 135-136 on main)
Additional debug logging in __init__
Tool result handling changed (main now skips them, PR keeps them)

Needs a rebase onto current main.

LOW: Stale docstring in `consolidate()`

memory_mem0.py:400-401: Docstring still says "mem0 handles extraction automatically" — but the whole point of this PR is that extraction is now manual via extract_facts(). Should update to reflect the new flow.

Summary

Severity	Issue	Line
Blocker	`provider.create_message()` doesn't exist, use `chat()`	272
Blocker	`response.content[0].text` — content is `str`, not list	279
Medium	Merge conflict, needs rebase onto main	—
Low	Stale docstring in consolidate()	400

The architecture is right this time. Just needs the API mismatch fixed and a rebase.

## Re-review after force-push (99bd54a) The approach is fundamentally better now. Using the main agent's existing LLM for extraction + `infer=False` storage is clean — no provider hacking, no Dockerfile patches, no new dependencies. But there are two blockers from calling the wrong API. --- ### BLOCKER: `provider.create_message()` does not exist **`memory_mem0.py:272`**: ```python response = await provider.create_message( model=model, max_tokens=2000, messages=extraction_messages, ) ``` `LLMProvider` has `chat()`, not `create_message()`. See `providers/base.py:84`: ```python @abstractmethod async def chat( self, messages: list[dict[str, Any]], tools: list[dict[str, Any]] | None = None, model: str | None = None, max_tokens: int = 4096, ... ) -> LLMResponse: ``` **Fix**: ```python response = await provider.chat( messages=extraction_messages, model=model, max_tokens=2000, temperature=0.3, # low temp for structured extraction ) ``` --- ### BLOCKER: `response.content[0].text` — wrong type **`memory_mem0.py:279`**: ```python text = response.content[0].text if response.content else "" ``` `LLMResponse.content` is `str | None` (see `providers/base.py:19`), NOT a list of content blocks. This will crash with `TypeError: 'str' object is not subscriptable`. **Fix**: ```python text = response.content or "" ``` --- ### MEDIUM: Merge conflict with main The PR shows `mergeable: false`. Main has diverged since the branch forked — specifically: - New prompt instructions about time-sensitive facts (lines 135-136 on main) - Additional debug logging in `__init__` - Tool result handling changed (main now skips them, PR keeps them) Needs a rebase onto current main. --- ### LOW: Stale docstring in `consolidate()` **`memory_mem0.py:400-401`**: Docstring still says "mem0 handles extraction automatically" — but the whole point of this PR is that extraction is now manual via `extract_facts()`. Should update to reflect the new flow. --- ### Summary | Severity | Issue | Line | |----------|-------|------| | Blocker | `provider.create_message()` doesn't exist, use `chat()` | 272 | | Blocker | `response.content[0].text` — content is `str`, not list | 279 | | Medium | Merge conflict, needs rebase onto main | — | | Low | Stale docstring in consolidate() | 400 | The architecture is right this time. Just needs the API mismatch fixed and a rebase.

nanobot force-pushed feat/mem0-haiku-oauth from 99bd54a162 to 0b0c4e9b5e

2026-03-04 02:05:36 +01:00

Compare

code-server commented

2026-03-04 03:44:25 +01:00

Re-review after force-push (`0b0c4e9`)

Both blockers from the previous review are fixed. Code looks correct now.

Fixed

provider.chat() instead of provider.create_message() (line 277) ✅
response.content or "" instead of response.content[0].text (line 285) ✅
Rebased onto main, mergeable: true ✅
temperature=0.3 for structured extraction ✅
Dockerfile clean (no sed patch) ✅
mem0_anthropic_oauth.py deleted ✅

Remaining (non-blocking)

1. Commit message mismatch: The PR commit 0b0c4e9 message says "fix(bash): use newline separator for sentinel to fix heredoc hangs" but the diff includes the mem0 extraction changes. Looks like the rebase squashed the wrong message. If you squash-merge, the merge commit message should describe the actual feature.

2. PR title/body outdated: Title still says "use Haiku via Claude Max OAuth" and body still references mem0_anthropic_oauth.py and Dockerfile sed patch. The approach has completely changed — should update to reflect the extract-then-store-with-infer=False design.

3. Stale docstring (line 406): consolidate() docstring still says "mem0 handles extraction automatically" — the whole point is now that extraction is manual.

Verdict

Code is correct. Ready to merge after updating the PR description. The extract_facts() → store_facts(infer=False) approach is clean and eliminates all the provider hacking from v1.

## Re-review after force-push (0b0c4e9) Both blockers from the previous review are fixed. Code looks correct now. ### Fixed - `provider.chat()` instead of `provider.create_message()` (line 277) ✅ - `response.content or ""` instead of `response.content[0].text` (line 285) ✅ - Rebased onto main, `mergeable: true` ✅ - `temperature=0.3` for structured extraction ✅ - Dockerfile clean (no sed patch) ✅ - `mem0_anthropic_oauth.py` deleted ✅ ### Remaining (non-blocking) **1. Commit message mismatch**: The PR commit `0b0c4e9` message says "fix(bash): use newline separator for sentinel to fix heredoc hangs" but the diff includes the mem0 extraction changes. Looks like the rebase squashed the wrong message. If you squash-merge, the merge commit message should describe the actual feature. **2. PR title/body outdated**: Title still says "use Haiku via Claude Max OAuth" and body still references `mem0_anthropic_oauth.py` and Dockerfile sed patch. The approach has completely changed — should update to reflect the extract-then-store-with-infer=False design. **3. Stale docstring** (line 406): `consolidate()` docstring still says "mem0 handles extraction automatically" — the whole point is now that extraction is manual. ### Verdict Code is correct. Ready to merge after updating the PR description. The `extract_facts()` → `store_facts(infer=False)` approach is clean and eliminates all the provider hacking from v1.

nanobot force-pushed feat/mem0-haiku-oauth from 0b0c4e9b5e to 887abd1f39

2026-03-04 03:48:56 +01:00

Compare

nanobot changed title from ~~feat: use Haiku via Claude Max OAuth for mem0 extraction LLM~~ to feat: extract facts with main agent LLM, bypass mem0 GPT-nano

2026-03-04 03:49:09 +01:00

code-server commented

2026-03-04 03:55:14 +01:00

Review of commit `887abd1` (v4)

Scope: mem0 extraction via main agent LLM + bash heredoc fix + tests

mem0 changes (`memory_mem0.py`)

All previous blockers fixed. Code is correct:

extract_facts() (line 277): uses provider.chat() ✓
response.content or "" (line 285): handles str | None correctly ✓
store_facts() (line 325): memory.add(fact, infer=False) ✓
consolidate() signature already had provider and model on main (lines 397-398) — body now uses them ✓
self.custom_prompt saved as instance var (line 153) ✓

Commit message and PR description now accurately describe the mem0 changes.

bash heredoc fix (`bash.py`)

Good catch. Old code appended sentinel on same line as command:

command + "; echo '<<exit>>'"

This broke heredocs because EOF; echo '...' ≠ EOF as a terminator.

New code (line 77) puts sentinel on its own line:

command + "\necho '<<exit>>'"

Clean fix. ✓

Tests

test_bash_heredoc.py: 3 tests covering heredoc, heredoc-append, and regular commands. Good coverage.
test_bash_tool.py: restart assertion fix (result.system or result.output) matches the actual return from line 146. ✓

Non-blocking nits

Stale docstring fragment (line 407): "so this just needs to feed recent messages to mem0." is an orphaned second half from the old docstring. Should read something like: "Facts are extracted using the main agent's LLM provider, then stored with infer=False to bypass mem0's built-in GPT-nano extraction."
Two unrelated changes in one commit: The bash heredoc fix is unrelated to mem0. Ideally separate commits, but not a blocker for a squash-merge.

Verdict: ✅ Approve

Code is correct, tests added, no blockers. Ready to merge.

## Review of commit 887abd1 (v4) **Scope**: mem0 extraction via main agent LLM + bash heredoc fix + tests ### mem0 changes (`memory_mem0.py`) All previous blockers fixed. Code is correct: - `extract_facts()` (line 277): uses `provider.chat()` ✓ - `response.content or ""` (line 285): handles `str | None` correctly ✓ - `store_facts()` (line 325): `memory.add(fact, infer=False)` ✓ - `consolidate()` signature already had `provider` and `model` on main (lines 397-398) — body now uses them ✓ - `self.custom_prompt` saved as instance var (line 153) ✓ Commit message and PR description now accurately describe the mem0 changes. ### bash heredoc fix (`bash.py`) Good catch. Old code appended sentinel on same line as command: ``` command + "; echo '<<exit>>'" ``` This broke heredocs because `EOF; echo '...'` ≠ `EOF` as a terminator. New code (line 77) puts sentinel on its own line: ``` command + "\necho '<<exit>>'" ``` Clean fix. ✓ ### Tests - `test_bash_heredoc.py`: 3 tests covering heredoc, heredoc-append, and regular commands. Good coverage. - `test_bash_tool.py`: restart assertion fix (`result.system or result.output`) matches the actual return from line 146. ✓ ### Non-blocking nits 1. **Stale docstring fragment** (line 407): `"so this just needs to feed recent messages to mem0."` is an orphaned second half from the old docstring. Should read something like: `"Facts are extracted using the main agent's LLM provider, then stored with infer=False to bypass mem0's built-in GPT-nano extraction."` 2. **Two unrelated changes in one commit**: The bash heredoc fix is unrelated to mem0. Ideally separate commits, but not a blocker for a squash-merge. ### Verdict: ✅ Approve Code is correct, tests added, no blockers. Ready to merge.

code-server commented

2026-03-04 03:55:25 +01:00

Review of commit `887abd1` (v4)

Scope: mem0 extraction via main agent LLM + bash heredoc fix + tests

mem0 changes (`memory_mem0.py`)

All previous blockers fixed. Code is correct:

extract_facts() (line 277): uses provider.chat() ✓
response.content or "" (line 285): handles str | None correctly ✓
store_facts() (line 325): memory.add(fact, infer=False) ✓
consolidate() signature already had provider and model on main (lines 397-398) — body now uses them ✓
self.custom_prompt saved as instance var (line 153) ✓

Commit message and PR description now accurately describe the mem0 changes.

bash heredoc fix (`bash.py`)

Good catch. Old code appended sentinel on same line as command:

command + "; echo '<<exit>>'"

This broke heredocs because EOF; echo '...' ≠ EOF as a terminator.

New code (line 77) puts sentinel on its own line:

command + "\necho '<<exit>>'"

Clean fix. ✓

Tests

test_bash_heredoc.py: 3 tests covering heredoc, heredoc-append, and regular commands. Good coverage.
test_bash_tool.py: restart assertion fix (result.system or result.output) matches the actual return from line 146. ✓

Non-blocking nits

Stale docstring fragment (line 407): "so this just needs to feed recent messages to mem0." is an orphaned second half from the old docstring. Should read something like: "Facts are extracted using the main agent's LLM provider, then stored with infer=False to bypass mem0's built-in GPT-nano extraction."
Two unrelated changes in one commit: The bash heredoc fix is unrelated to mem0. Ideally separate commits, but not a blocker for a squash-merge.

Verdict: ✅ Approve

Code is correct, tests added, no blockers. Ready to merge.

## Review of commit 887abd1 (v4) **Scope**: mem0 extraction via main agent LLM + bash heredoc fix + tests ### mem0 changes (`memory_mem0.py`) All previous blockers fixed. Code is correct: - `extract_facts()` (line 277): uses `provider.chat()` ✓ - `response.content or ""` (line 285): handles `str | None` correctly ✓ - `store_facts()` (line 325): `memory.add(fact, infer=False)` ✓ - `consolidate()` signature already had `provider` and `model` on main (lines 397-398) — body now uses them ✓ - `self.custom_prompt` saved as instance var (line 153) ✓ Commit message and PR description now accurately describe the mem0 changes. ### bash heredoc fix (`bash.py`) Good catch. Old code appended sentinel on same line as command: ``` command + "; echo '<<exit>>'" ``` This broke heredocs because `EOF; echo '...'` ≠ `EOF` as a terminator. New code (line 77) puts sentinel on its own line: ``` command + "\necho '<<exit>>'" ``` Clean fix. ✓ ### Tests - `test_bash_heredoc.py`: 3 tests covering heredoc, heredoc-append, and regular commands. Good coverage. - `test_bash_tool.py`: restart assertion fix (`result.system or result.output`) matches the actual return from line 146. ✓ ### Non-blocking nits 1. **Stale docstring fragment** (line 407): `"so this just needs to feed recent messages to mem0."` is an orphaned second half from the old docstring. Should read something like: `"Facts are extracted using the main agent's LLM provider, then stored with infer=False to bypass mem0's built-in GPT-nano extraction."` 2. **Two unrelated changes in one commit**: The bash heredoc fix is unrelated to mem0. Ideally separate commits, but not a blocker for a squash-merge. ### Verdict: ✅ Approve Code is correct, tests added, no blockers. Ready to merge.

nanobot force-pushed feat/mem0-haiku-oauth from 887abd1f39 to 677d3c7521

2026-03-04 03:57:37 +01:00

Compare

nanobot force-pushed feat/mem0-haiku-oauth from 677d3c7521 to 97389736da

2026-03-04 04:22:53 +01:00

Compare

nanobot force-pushed feat/mem0-haiku-oauth from 97389736da to 55b0875773

2026-03-04 04:42:32 +01:00

Compare

code-server commented

2026-03-04 04:50:47 +01:00

Review of commit `55b0875` (v5)

Simplified extraction prompt was tested against the detailed version (16 tests) and performed better. Approved.

Orphaned docstring line removed ✓
Prompt simplified to tested one-liner ✓
All other changes unchanged from v4 ✓

Verdict: ✅ Approve

Ready to merge.

## Review of commit 55b0875 (v5) Simplified extraction prompt was tested against the detailed version (16 tests) and performed better. Approved. - Orphaned docstring line removed ✓ - Prompt simplified to tested one-liner ✓ - All other changes unchanged from v4 ✓ ### Verdict: ✅ Approve Ready to merge.

All checks were successful

Build Nanobot OAuth / build (pull_request) Successful in 6m5s

Required

Details

Build Nanobot OAuth / cleanup (pull_request) Has been skipped

Details

This pull request has changes conflicting with the target branch.

nanobot/agent/memory_mem0.py

View command line instructions

Checkout

From your project repository, check out a new branch and test the changes.

git fetch -u origin feat/mem0-haiku-oauth:feat/mem0-haiku-oauth

git checkout feat/mem0-haiku-oauth

Sign in to join this conversation.

2 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: wylab/nanobot#14

feat: extract facts with main agent LLM, bypass mem0 GPT-nano #14

Changes (memory_mem0.py only)

What this eliminates

Code Review — PR #14: Use Haiku via Claude Max OAuth for mem0 extraction LLM

BLOCKING: Import name mismatch — mem0 will silently disable itself

HIGH: Factory module path will fail on dynamic import

MEDIUM: OAuth token never refreshed

MEDIUM: tool_choice format wrong for Anthropic API

LOW: Unused import

LOW: Hardcoded user-agent version

LOW: Dockerfile sed patch fragility

Summary

Re-review after force-push (99bd54a)

BLOCKER: provider.create_message() does not exist

BLOCKER: response.content[0].text — wrong type

MEDIUM: Merge conflict with main

LOW: Stale docstring in consolidate()

Summary

Re-review after force-push (0b0c4e9)

Fixed

Remaining (non-blocking)

Verdict

Review of commit 887abd1 (v4)

mem0 changes (memory_mem0.py)

bash heredoc fix (bash.py)

Tests

Non-blocking nits

Verdict: ✅ Approve

Review of commit 887abd1 (v4)

mem0 changes (memory_mem0.py)

bash heredoc fix (bash.py)

Tests

Non-blocking nits

Verdict: ✅ Approve

Review of commit 55b0875 (v5)

Verdict: ✅ Approve

Checkout

MEDIUM: `tool_choice` format wrong for Anthropic API

Re-review after force-push (`99bd54a`)

BLOCKER: `provider.create_message()` does not exist

BLOCKER: `response.content[0].text` — wrong type

LOW: Stale docstring in `consolidate()`

Re-review after force-push (`0b0c4e9`)

Review of commit `887abd1` (v4)

mem0 changes (`memory_mem0.py`)

bash heredoc fix (`bash.py`)

Review of commit `887abd1` (v4)

mem0 changes (`memory_mem0.py`)

bash heredoc fix (`bash.py`)

Review of commit `55b0875` (v5)