keep-v0.1.0-alpha.3

Keep 0.1.0-alpha.3 release notes

Keep provides shared local memory for AI agents through stdio or protected HTTP.
This alpha provides Linux x64 with glibc, Windows x64, and macOS Apple Silicon archives.
The Linux archive requires glibc 2.39 or later, such as Ubuntu 24.04.

HTTP mode now accepts --bind-address with a literal IPv4 or IPv6 address.
Use it with --port to select an interface or a container listener. The default remains 127.0.0.1.
Bearer authentication, Origin rejection, and exact Host checks remain required.
The HTTP reference explains listener addresses and HTTPS proxies.
The document API, license, and dependency notices remain unchanged.
Windows includes the required Microsoft Visual C++ runtime.

Tested systems

Condition	Linux x64	Windows x64	macOS arm64
OS	Ubuntu 24.04.5 LTS; kernel 6.8.0-142-generic	Windows 11, build 10.0.26200	macOS 26.5.1
Visible logical CPUs	16	32	8
Physical memory bytes	16,765,132,800	66,217,689,088	17,179,869,184
Model, cache recovery, and backup checks	Passed	Passed	Passed

These are tested hosts, not minimum system requirements. No clean-machine installation result is claimed.

Model and limits

Default search is lexical and needs no model. Add --semantic to enable hybrid search with a local model.
The archives contain no model weights; the first enabled start can download about 86 MiB.
The pinned model is Snowflake/snowflake-arctic-embed-xs, revision d8c86521100d3556476a063fc2342036d45c106f.
All three platforms passed model checksum checks, corrupt-cache recovery, hybrid search, and lexical fallback after model failure.

Hybrid search permits at most 512 documents, 4,096 chunks, and 8,388,608 searchable text bytes across the store.
Narrow filters do not bypass these limits. Use lexical search when a model is unavailable or a hybrid limit is exceeded.
The measured dataset contains 512 documents, 4,096 chunks, and 6,814,994 searchable text bytes.
It does not exercise the complete 8 MiB text limit. Lexical search remained available after a 513th document exceeded the document limit.

Observed costs

Measurement	Linux x64	Windows x64	macOS arm64
Initial download, verification, and model load	2.320 s	6.656 s	2.478 s
Model load with verified cache in a fresh process	111.277 ms	79.000 ms	53.427 ms
Index wall time	770.966 s	648.750 s	574.062 s
Index CPU time	1,006.890 s	3,659.578 s	1,279.270 s
Sampled peak resident memory	206,565,376 bytes	215,281,664 bytes	207,683,584 bytes
One warm hybrid query	194.334 ms	469.000 ms	465.121 ms
Maximum sample gap during initial download and load	46.846 ms	657.000 ms	418.003 ms
Maximum sample gap during cached-process measurements	78.848 ms	79.000 ms	439.001 ms

The hosts are shared, with no exclusive CPU allocation. Linux has a two-CPU quota and a 4 GiB memory limit.
Inherited Linux limits can be stricter. Windows job-object restrictions were not queried.

Cached measurements use a fresh process, verified model files, and the same dataset. The OS file cache stays warm.
Model and index times use observed phase transitions. Index CPU time includes protocol work; resident memory is a sampled peak.
The warm query cost covers one request after readiness.

The requested sample interval is 10 ms. Actual gaps limit precision.
Treat model-load intervals shorter than their maximum sample gaps as coarse observations.
These results establish no general speed benefit, memory maximum, or comparison with another product.

Fresh-agent tasks

Fresh agents completed all five tasks on each native archive.
The tasks covered exact queries, conditional edits, history restoration, archive access, and deletion.
Linux used 22 MCP calls, Windows used 24, and macOS used 20.
Each trial had four requested mutations and one expected not_found response. No unexpected MCP errors occurred.
The evaluator model was GPT-6.1 Sol with medium reasoning. These trials provide exploratory evidence without a product comparison.
Session times include interface readiness and bridge delays. The Windows start time is approximate; client token counts are unavailable.

Integrity and recovery

This alpha uses checksums-only integrity. The archives are unsigned.
Download the archive and SHA256SUMS from the same prerelease. Compare the archive's SHA-256 before extraction.
Use the public README for installation and client setup.
Stop Keep before a backup or restore. A restored backup loses later writes.
Uninstall removes only the executable; the personal store and backups remain on disk.

keep-v0.1.0-alpha.2

Keep 0.1.0-alpha.2 release notes

Keep provides shared local memory for AI agents through stdio or protected HTTP.
This alpha provides Linux x64 with glibc, Windows x64, and macOS Apple Silicon archives.

Windows now includes the MSVC runtime in the executable. No separate Visual C++ Redistributable installation is required for alpha.2.
Model downloads no longer fail intermittently when a server closes its connection after each file.
The document API, license, and dependency notices remain unchanged.

Tested systems

Condition	Linux x64	Windows x64	macOS arm64
OS	Ubuntu 24.04.5 LTS; kernel 6.8.0-142-generic	Windows 11, build 10.0.26200	macOS 26.5.1
Visible logical CPUs	16	32	8
Physical memory bytes	16,765,132,800	66,217,689,088	17,179,869,184
Model, cache recovery, and backup checks	Passed	Passed	Passed

These are tested hosts, not minimum system requirements. No clean-machine installation result is claimed.

Model and limits

Default search is lexical and needs no model. Add --semantic to enable hybrid search with a local model.
The archives contain no model weights; the first enabled start can download about 86 MiB.
The pinned model is Snowflake/snowflake-arctic-embed-xs, revision d8c86521100d3556476a063fc2342036d45c106f.
All three platforms passed model checksum checks, corrupt-cache recovery, hybrid search, and lexical fallback after model failure.

Hybrid search permits at most 512 documents, 4,096 chunks, and 8,388,608 searchable text bytes across the store.
Narrow filters do not bypass these limits. Use lexical search when a model is unavailable or a hybrid limit is exceeded.
The measured dataset contains 512 documents, 4,096 chunks, and 6,814,994 searchable text bytes.
It does not exercise the complete 8 MiB text limit. Lexical search remained available after a 513th document exceeded the document limit.

Observed costs

Measurement	Linux x64	Windows x64	macOS arm64
Initial download, verification, and model load	2.587 s	6.625 s	3.295 s
Model load with verified cache in a fresh process	97.475 ms	62.000 ms	60.829 ms
Index wall time	842.655 s	681.516 s	615.514 s
Index CPU time	1,090.380 s	3,571.594 s	1,427.570 s
Sampled peak resident memory	208,609,280 bytes	219,582,464 bytes	227,229,696 bytes
One warm hybrid query	321.129 ms	281.000 ms	350.763 ms
Maximum sample gap during initial download and load	31.123 ms	3,546.000 ms	462.880 ms
Maximum sample gap during cached-process measurements	33.148 ms	63.000 ms	403.011 ms

The hosts are shared, with no exclusive CPU allocation. Linux has a two-CPU quota and a 4 GiB memory limit.
Inherited Linux limits can be stricter. Windows job-object restrictions were not queried.

Cached measurements use a fresh process, verified model files, and the same dataset. The OS file cache stays warm.
Model and index times use observed phase transitions. Index CPU time includes protocol work; resident memory is a sampled peak.
The warm query cost covers one request after readiness.

The requested sample interval is 10 ms. Actual gaps limit precision.
The Windows initial-load observation spans 6.625 seconds, with a 3.546-second maximum sample gap; it is coarse.
Windows and macOS cached model-load intervals are shorter than their maximum sample gaps and are also coarse.
These results establish no general speed benefit, memory maximum, or comparison with another product.

Fresh-agent tasks

Fresh agents completed all five tasks on each native archive.
The tasks covered exact queries, conditional edits, history restoration, archive access, and deletion.
Linux and Windows used 19 MCP calls; macOS used 20.
Each trial had four requested mutations and one expected not_found response. No unexpected MCP errors occurred.
The evaluator model was GPT-6.1 Sol with medium reasoning. These trials provide exploratory evidence without a product comparison.

Integrity and recovery

This alpha uses checksums-only integrity. The archives are unsigned.
Download the archive and SHA256SUMS from the same prerelease. Compare the archive's SHA-256 before extraction.
Use the public README for installation and client setup.
Stop Keep before a backup or restore. A restored backup loses later writes.
Uninstall removes only the executable; the personal store and backups remain on disk.

keep-v0.1.0-alpha.1

Keep 0.1.0-alpha.1 release notes

Keep stores personal text and JSON documents locally for MCP clients.
The alpha provides Linux x64 GNU, Windows x64, and macOS Apple Silicon archives.
It exposes only the document API through stdio or protected local HTTP. Default search is lexical.

Tested systems

Condition	Linux x64 GNU	Windows x64	macOS arm64
OS	Ubuntu 24.04.5 LTS; kernel 6.8.0-139-generic	Windows 11, build 10.0.26200	macOS 26.5.1
Visible logical CPUs	16	32	8
Physical memory bytes	16,765,136,896	66,217,689,088	17,179,869,184
Native model, recovery, and backup checks	Passed	Passed	Passed

Windows requires the Microsoft Visual C++ v14 Redistributable for x64. The archive contains no runtime DLLs.
These are tested hosts, not minimum system requirements. No clean-machine installation result is claimed.

Model and limits

The optional --semantic flag enables hybrid search. The archives contain no model weights.
The pinned model is Snowflake/snowflake-arctic-embed-xs, revision d8c86521100d3556476a063fc2342036d45c106f.
Each native probe verified the three model files against their pinned SHA-256 checksums.
Each probe rejected corrupt model bytes, repaired the cache, reached hybrid readiness, and returned a hybrid result.
Lexical search recovered after model failure and after a 513th document exceeded the hybrid document budget.

The synthetic recipe, workout, household, and prose fixture has 512 documents, 4,096 chunks, and 6,814,994 searchable text bytes.
The text size is approximately 6.499 MiB. It does not test the complete 8 MiB text boundary.
Hybrid search permits at most 512 documents, 4,096 chunks, and 8,388,608 searchable text bytes across the store.
Narrow filters do not bypass these index limits. Use lexical mode when a model is unavailable or a hybrid budget is exceeded.

Observed costs

Measurement	Linux x64 GNU	Windows x64	macOS arm64
Initial provision, verification, and model load	7.143 s	5.610 s	5.513 s
Model load with verified cache in a fresh process	44.813 ms	63.000 ms	54.533 ms
Index wall time	658.378 s	730.969 s	480.016 s
Index CPU time	865.620 s	3,707.734 s	1,108.360 s
Sampled peak resident memory	207,962,112 bytes	219,881,472 bytes	273,842,176 bytes
One warm hybrid query	200.781 ms	266.000 ms	213.165 ms
Maximum sample gap during initial provision	55.316 ms	94.000 ms	170.191 ms
Maximum sample gap during cached-process measurements	28.033 ms	63.000 ms	563.357 ms

The hosts are shared and have no exclusive CPU allocation. Other workloads can affect the observations.
Linux exposes a two-CPU container quota and a 4 GiB memory limit despite 16 visible logical CPUs.
Inherited container limits can be stricter. Windows job-object restrictions were not queried.
Initial provision includes the download, checksum verification, and model load through the first observed index phase.
Cached measurements use a fresh process, verified model files, and the same fixture. The OS file cache stays warm.
Index wall time spans observed index preparation through readiness. Index CPU includes the child process’s protocol work.
Resident memory is a sampled peak. The warm query cost covers one request after readiness.

The requested sample interval is 10 ms. The maximum observed gaps limit phase timing precision.
The macOS 54.533 ms model-load observation is shorter than its 563.357 ms maximum gap; it is a coarse observed interval.
These results establish no general speed benefit, memory maximum, or comparison with another product.

Fresh-agent tasks

Target	Independently scored result
Linux x64	Passed all five tasks; 20 evaluator tool calls; zero unexpected MCP errors
macOS arm64	Passed all five tasks; 27 evaluator tool calls; zero unexpected MCP errors
Windows x64	Passed all five tasks; 22 evaluator tool calls; zero unexpected MCP errors

The tasks cover recipe sources, complete workout enumeration, a plan correction and restoration, household records, archive access, and deletion.
All three trials found four walks totaling 15 km and preserved the plan’s budget and source.
The Linux verification read after deletion returned the expected not_found error.
The evaluator configuration was GPT-6 Luna with medium reasoning. The backend model identifier was unavailable.
The macOS trial used the exact production archive locally on the native build host after an Actions setup failure before binary execution.
The coordinator supplied a readiness command with a stray argument; both evaluators removed it.
Linux corrected a local stop-request literal. macOS repeated a read-only resource request after reported missing output.
Windows repeated schema discovery and a read-only cursor page after reported output/display issues.
These reports do not establish a server or transport failure cause. No duplicate mutation occurred.

Linux reported 480 seconds. Windows reported 510 seconds, including interface and schema discovery.
macOS reported 551 seconds after initial readiness, excluding initial discovery and recovery.
These intervals do not measure complete installation or comparable first-use duration. Token counts were unavailable.

Integrity and recovery

This alpha uses checksums-only integrity. The archives are unsigned.
Download the archive and SHA256SUMS from the same prerelease. Compare the archive’s SHA-256 before extraction.
Use the public README for installation, client setup, backup, and recovery.
Stop Keep before a backup or restore. A restored backup loses later writes.
Uninstall removes only the executable; the personal store and backups remain on disk.