Skip to content

README.md

Latest commit

 

History

History
139 lines (103 loc) · 5.9 KB

README.md

File metadata and controls

139 lines (103 loc) · 5.9 KB

llama-index-tools-goodmem

GoodMem is a server-side memory layer for AI agents with semantic storage, retrieval, and LLM-powered summarization. This package exposes GoodMem's API as a LlamaIndex BaseToolSpec, so any LlamaIndex agent can store, search, and manage long-term memories.

Install

pip install llamaindex-goodmem

The PyPI distribution is llamaindex-goodmem; the Python import path is llama_index.tools.goodmem (the llama_index.* namespace is shared with the LlamaIndex ecosystem).

For local development from this repo:

pip install -e .            # editable install
pip install -e ".[dev]"     # editable install + test deps

Quick start

from llama_index.tools.goodmem import GoodMemToolSpec

tool_spec = GoodMemToolSpec(
    api_key="gm_xxx",
    base_url="https://api.goodmem.ai",
    verify_ssl=True,    # set False for a self-signed dev cert
)

# Use with a LlamaIndex agent
tools = tool_spec.to_tool_list()

Constructor parameters

Parameter Type Required Description
api_key str yes GoodMem API key (sent as X-API-Key)
base_url str yes Base URL of your GoodMem server
verify_ssl bool no Verify TLS certificates. Default True.
timeout float no Per-request timeout in seconds. Default 120 (LLM-summary retrievals can take tens of seconds).

Tools

The tool spec exposes 11 sync/async tool pairs. The names below are the sync entry points; each has an async counterpart prefixed with a (e.g. aretrieve_memories).

Tool Purpose
list_embedders List server-managed embedder models.
list_spaces List spaces.
get_space Fetch one space by ID.
create_space Create a new space (or reuse one with the same name).
update_space Rename a space, toggle publicRead, or merge/replace labels.
delete_space Delete a space (cascades to its memories).
create_memory Store a text or file payload as a memory.
list_memories Paginate memories in a space (with optional status / metadata filters).
retrieve_memories Semantic search; optional rerank, threshold, LLM summary, chrono-resort.
get_memory Fetch one memory and (optionally) its original content.
delete_memory Delete a memory.

retrieve_memories returns List[llama_index.core.schema.Document]. Each Document.text is a matched chunk; Document.metadata includes chunkId, memoryId, relevanceScore, resultSetId, query, plus abstractReply if an llm_id was passed.

Post-processor knobs (retrieve_memories)

Argument Type Notes
reranker_id UUID Reranks results by direct query–chunk scoring.
llm_id UUID Generates an abstractReply summary across the result set.
relevance_threshold float 0..1 Drops results below this score.
llm_temperature float 0..2 LLM creativity (sent on the wire as llm_temp).
chronological_resort bool Re-sort the post-processor output by creation time.

Examples

# Create a space
space = tool_spec.create_space(name="my-research", embedder_id="<uuid>")

# Add memories
tool_spec.create_memory(space_id=space["spaceId"], text_content="...")
tool_spec.create_memory(space_id=space["spaceId"], file_path="/path/to/doc.pdf")

# Search
docs = tool_spec.retrieve_memories(
    query="What's the conclusion?",
    space_ids=[space["spaceId"]],
    max_results=5,
    llm_id="<llm-uuid>",
    llm_temperature=0.2,
)
for d in docs:
    print(d.metadata["relevanceScore"], d.text[:120])

# LLM summary (when llm_id is set)
print(docs[0].metadata["abstractReply"]["text"])

Use inside a LlamaIndex agent

from llama_index.tools.goodmem import GoodMemToolSpec
from llama_index.core.agent import ReActAgent
from llama_index.llms.openai import OpenAI

tool_spec = GoodMemToolSpec(api_key="gm_xxx", base_url="https://api.goodmem.ai")
agent = ReActAgent.from_tools(tool_spec.to_tool_list(), llm=OpenAI(model="gpt-4o-mini"))

response = agent.chat("Remember that the project deadline is May 14.")

Testing

Unit tests (no live server)

pip install -e ".[dev]"
pytest tests/test_tools_goodmem.py -v

Live e2e smoke test

Exercises every one of the 11 tools and every post-processor knob against a running GoodMem server. Mirrors the §7 smoke-test plan in the GoodMem build guide.

GOODMEM_API_KEY=gm_xxx \
GOODMEM_BASE_URL=https://localhost:8080 \
GOODMEM_EMBEDDER_ID=<embedder-uuid> \
GOODMEM_RERANKER_ID=<reranker-uuid> \
GOODMEM_LLM_ID=<llm-uuid> \
GOODMEM_VERIFY_SSL=false \
pytest tests/test_tools_goodmem_e2e.py -v

GOODMEM_RERANKER_ID and GOODMEM_LLM_ID are optional — the variants that need them are auto-skipped if unset. The full 18-step smoke test passes when all three are configured.

License

MIT