aether-context 0.3.1__tar.gz → 0.5.0__tar.gz

{aether_context-0.3.1/aether_context.egg-info → aether_context-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: aether-context
-Version: 0.3.1
+Version: 0.5.0
 Summary: Unlimited Context — virtual memory for an LLM's attention. Local-first, numpy-only core.
 Author: Aether AI
 Maintainer: Aether AI
@@ -34,6 +34,9 @@ Requires-Dist: transformers>=4.40; extra == "hf"
 Requires-Dist: torch>=2.2; extra == "hf"
 Provides-Extra: fast
 Requires-Dist: hnswlib>=0.8; extra == "fast"
+Provides-Extra: hosted
+Requires-Dist: pydantic<3,>=2.10; extra == "hosted"
+Requires-Dist: cryptography<51,>=44; extra == "hosted"
 Provides-Extra: all
 Requires-Dist: aether-context[fast,hf,llamacpp]; extra == "all"
 Provides-Extra: dev
@@ -42,6 +45,7 @@ Requires-Dist: pytest-cov; extra == "dev"
 Requires-Dist: ruff>=0.5; extra == "dev"
 Requires-Dist: mypy>=1.10; extra == "dev"
 Requires-Dist: pyyaml>=6; extra == "dev"
+Requires-Dist: aether-context[hosted]; extra == "dev"
 Dynamic: license-file
 
 <div align="center">
@@ -86,6 +90,27 @@ aether-context setup
 `setup` sizes the pool, checks for a local model, and verifies the engine end to end. It works
 with no daemon, no network and no model pulled — the check runs against the built-in mock model.
 
+Two `run` commands can use the same pool directory. The first task is stored as user-authored
+memory; the second retrieves it before generation, even with the default `separate` pool mode:
+
+```bash
+aether-context run "remember: the deploy key rotates on Fridays" --dir ./pool
+aether-context run "when does the deploy key rotate?" --dir ./pool
+```
+
+With the default mock model, the status lines from a fresh pool are:
+
+```text
+[pool 5 GB | slices 2/2275390 | reach 1.17B tok | hit 0% | recalled 0 | mode separate | index flat]
+[pool 5 GB | slices 4/2275390 | reach 1.17B tok | hit 0% | recalled 1 | mode separate | index flat]
+```
+
+The mock prints synthetic text, so use a real local model for an answer to the question.
+`recalled` counts slices supplied before generation; `hit` remains the pager's warm-cache rate.
+Only user and tool memory is supplied by default. `run --include-model-memory` also supplies
+model-authored spill. Retrieved context is capped at 20% of the model window, with room reserved
+for the task and output; a slice too large to fit is skipped.
+
 ```python
 from aether_context import Session
 
@@ -242,7 +267,9 @@ session its own pool and index — fully isolated and private, but you pay one i
 RAM scales with `N × pool` (that's the last column above). `--pool-mode shared` pays for the index
 **once**; each extra session adds only ~30 MB, so 50–70+ sessions fit and CPU becomes the limit
 instead of memory. The trade-off is that sessions can see each other's context. Use shared for
-related work on one project, separate for unrelated tasks.
+related work on one project, separate for unrelated tasks. A `run` invocation is a special case:
+it recalls prior user/tool facts from the same pool directory on reopen, even in separate mode.
+Use different directories when those tasks must not share memory.
 
 **How much building is that?** A ~128K window fills after well under an hour of active agent work,
 then starts compacting and forgetting. Assuming a busy coding agent encodes ~300K–1M keep-worthy

{aether_context-0.3.1 → aether_context-0.5.0}/README.md

@@ -40,6 +40,27 @@ aether-context setup
 `setup` sizes the pool, checks for a local model, and verifies the engine end to end. It works
 with no daemon, no network and no model pulled — the check runs against the built-in mock model.
 
+Two `run` commands can use the same pool directory. The first task is stored as user-authored
+memory; the second retrieves it before generation, even with the default `separate` pool mode:
+
+```bash
+aether-context run "remember: the deploy key rotates on Fridays" --dir ./pool
+aether-context run "when does the deploy key rotate?" --dir ./pool
+```
+
+With the default mock model, the status lines from a fresh pool are:
+
+```text
+[pool 5 GB | slices 2/2275390 | reach 1.17B tok | hit 0% | recalled 0 | mode separate | index flat]
+[pool 5 GB | slices 4/2275390 | reach 1.17B tok | hit 0% | recalled 1 | mode separate | index flat]
+```
+
+The mock prints synthetic text, so use a real local model for an answer to the question.
+`recalled` counts slices supplied before generation; `hit` remains the pager's warm-cache rate.
+Only user and tool memory is supplied by default. `run --include-model-memory` also supplies
+model-authored spill. Retrieved context is capped at 20% of the model window, with room reserved
+for the task and output; a slice too large to fit is skipped.
+
 ```python
 from aether_context import Session
 
@@ -196,7 +217,9 @@ session its own pool and index — fully isolated and private, but you pay one i
 RAM scales with `N × pool` (that's the last column above). `--pool-mode shared` pays for the index
 **once**; each extra session adds only ~30 MB, so 50–70+ sessions fit and CPU becomes the limit
 instead of memory. The trade-off is that sessions can see each other's context. Use shared for
-related work on one project, separate for unrelated tasks.
+related work on one project, separate for unrelated tasks. A `run` invocation is a special case:
+it recalls prior user/tool facts from the same pool directory on reopen, even in separate mode.
+Use different directories when those tasks must not share memory.
 
 **How much building is that?** A ~128K window fills after well under an hour of active agent work,
 then starts compacting and forgetting. Assuming a busy coding agent encodes ~300K–1M keep-worthy

{aether_context-0.3.1 → aether_context-0.5.0}/aether_context/__init__.py

@@ -24,6 +24,6 @@ Public surface (intentionally tiny):
 from aether_context.local_llm import load_model
 from aether_context.session import Session
 
-__version__ = "0.3.1"
+__version__ = "0.5.0"
 
 __all__ = ["Session", "load_model", "__version__"]

{aether_context-0.3.1 → aether_context-0.5.0}/aether_context/cli.py

@@ -158,6 +158,10 @@ def build_parser() -> argparse.ArgumentParser:
     )
     p_run.add_argument("task", type=str, help="the task/prompt to run.")
     _add_session_flags(p_run)
+    p_run.add_argument(
+        "--include-model-memory", action="store_true",
+        help="also supply model-authored spill as retrieved context (excluded by default).",
+    )
 
     p_chat = subparsers.add_parser(
         "chat", help="interactive REPL with slash-commands (/clear /new /status /quit ...)."
@@ -539,6 +543,7 @@ def _build_session(args: argparse.Namespace) -> Session:
         pool_index=args.index,
         pool_dir=pool_dir,
         mpo_chain=getattr(args, "mpo_chain", True),
+        include_model_memory=getattr(args, "include_model_memory", False),
         fallback_to_mock=True,
     )
 
@@ -564,6 +569,7 @@ def _status_line(s: dict[str, Any]) -> str:
     return (
         f"[pool {s['pool_gb']} GB | slices {s['slices_used']}/{s['capacity']} | "
         f"reach {int(s['reach_tokens']) / 1e9:.2f}B tok | hit {float(s['hit_rate']):.0%} | "
+        f"recalled {s['recalled']} | "
         f"mode {s['pool_mode']} | index {s['index']}]"
     )
 
@@ -885,6 +891,7 @@ def _status_lines(s: dict[str, Any]) -> list[str]:
         f"  slices:      {s['slices_used']} / {s['capacity']}",
         f"  reach:       {reach:,} tokens",
         f"  hit rate:    {float(s['hit_rate']):.0%}",
+        f"  recalled:    {s['recalled']} slices (last run)",
         f"  resident RAM ~= {s['resident_ram_mb']} MB (estimate)",
         f"  pool-mode:   {s['pool_mode']}",
         f"  index:       {s['index']}",

{aether_context-0.3.1 → aether_context-0.5.0}/aether_context/config.py

@@ -174,12 +174,14 @@ class SessionConfig:
     verbatim_fraction: float = VERBATIM_FRACTION
     trigger_fraction: float = TRIGGER_FRACTION
     target_fraction: float = TARGET_FRACTION
+    recall_fraction: float = 0.20
 
     def __post_init__(self) -> None:
         for name, value in (
             ("verbatim_fraction", self.verbatim_fraction),
             ("trigger_fraction", self.trigger_fraction),
             ("target_fraction", self.target_fraction),
+            ("recall_fraction", self.recall_fraction),
         ):
             if not (0.0 <= float(value) <= 1.0):
                 raise PoolBudgetError(